INDEPENDENT INTELLIGENCEAI AGENTS · AUTOMATION · WORK

THE AGENTIC CURRENT

Practical intelligence on AI agents, automation, and the future of work.

The Agentic Index

22 sourced profiles · Reviewed October 8, 2026

Know the system before you delegate

The Agentic Index is an editorial reference to real AI agents, automation platforms and the infrastructure behind them. Compare what a system is documented to do, where it operates, who grants access and what remains uncertain.

Inclusion and ordering are editorial. Placement is not a ranking, endorsement or paid recommendation. Product descriptions below reflect the linked vendors’ documentation; the buyer questions are The Agentic Current’s analysis. We have not independently tested these products.

Capabilities, product names, pricing and eligibility change quickly. These are dated research snapshots, not continuous monitoring. Verify consequential purchasing and deployment decisions with current documentation and your vendor agreement.

Choose a route

Choose a category or capability to jump to the relevant section, then expand a profile for details. If a product is absent, it is outside this first cohort; absence does not establish that it lacks agent capabilities.

How to read the evidence

Level 1 — Announced

Meaning: A vendor describes or announces the capability.

Does not establish: Reviewed sources do not establish access in a usable form.

Level 2 — Publicly available

Meaning: Documentation establishes a usable route, including paid access, open-source packages or a documented beta.

Does not establish: This does not mean every feature is generally available.

Level 3 — Documented deployment

Meaning: A specific real workflow or organization is documented. The entry must identify whether the account comes from a vendor, customer or independent reporter, and its scope.

Does not establish: A documented workflow does not establish suitability for every organization or workload.

Level 4 — Independently evaluated

Meaning: Relevant third-party testing or analysis supports a specified claim. The task, environment and limitations must be stated.

Does not establish: Evaluation of a specified claim does not establish suitability for every workload.

Levels describe the evidence assessed for this entry, not product quality, safety or a purchasing recommendation. Availability is a separate field. Vendor benchmarks do not automatically raise an evidence level, and a rubric score is not a production success rate. A higher level would still not establish suitability for your workload.

This first cohort uses official sources. Twenty-one entries are Level 2; ServiceNow is conservatively Level 1 because the reviewed product page alone does not establish feature-level access. We have not assigned deployment or independent-evaluation levels without reviewing that evidence. “Not clearly documented” means the reviewed sources did not settle the field; it does not mean the feature is absent.

Future sponsored or affiliate relationships, if any, will be disclosed separately and will not determine inclusion order or evidence ratings. See our Editorial Standards for attribution, analysis and corrections.

Browse by name

Assistants and research

Start here for delegated knowledge work. A research report, an editable deliverable and an authorized business-system action are different outputs.

ChatGPT Work

ChatGPT WorkGeneral-purpose assistantEvidence: Level 2 — Publicly available

OpenAI · Assistants and research
Also: General-purpose assistant; background execution

OpenAI describes Work as its environment for multi-step tasks that produce finished deliverables. It replaces the retired ChatGPT agent experience.

Availability: Generally available; plan, workspace and rollout restrictions apply.
Evidence: Level 2 — Publicly available
Primary users: Knowledge workers; consumers; business teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Task-based work across web, desktop and mobile; tools vary by experience.
  • Models: OpenAI models; selection and access vary by plan.
  • Tools and integrations: Connected apps and plugins can expose retrieval and actions.
  • Browser / computer use: Access depends on the selected environment and enabled tools.
  • API / agent building / multi-agent: Separate developer offerings; not treated as equivalent to Work.
  • Memory / persistence: Tasks and workspace context; cross-task memory scope not established here.
  • Approvals and permissions: App approval settings, provider permissions and workspace restrictions apply.
  • Auditability: Activity visibility varies; enterprise log-export coverage not clearly documented.
  • Deployment: OpenAI cloud and supported clients.
  • Pricing: Plan allowance shared with Codex; additional credits available.
  • Limitations: An app connection does not grant access beyond the connected account.

Sources & Further Reading
ChatGPT Work and Codex
Connected apps in ChatGPT
Managing app permissions
Retired ChatGPT agent experience

Buyer question — our analysis: Which actions can execute without review, and can administrators retrieve an adequate activity record?

Back to Index routes

Claude

ClaudeGeneral-purpose assistantEvidence: Level 2 — Publicly available

Anthropic · Assistants and research
Also: General-purpose assistant; research; connected work

Anthropic describes Claude as a work assistant that can produce editable deliverables and connect to tools. Cowork and chat are converging into one experience.

Availability: Generally available; new unified Cowork/chat experience is rolling out by plan.
Evidence: Level 2 — Publicly available
Primary users: Individuals; knowledge workers; business teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Chat and delegated work; background task support is plan-dependent.
  • Models: Anthropic Claude models; access varies by plan.
  • Tools and integrations: Skills, plugins, custom connectors and MCP integrations.
  • Browser / computer use: Documented for supported experiences; desktop can access selected folders and applications.
  • API / agent building / multi-agent: Separate Claude API; do not assume assistant features are API features.
  • Memory / persistence: Projects and memory features; retention and scope depend on settings.
  • Approvals and permissions: Selected access and organization controls; verify action-specific approval behavior.
  • Auditability: Complete action-log export coverage not clearly documented.
  • Deployment: Cloud; web, mobile and macOS/Windows clients.
  • Pricing: Free and subscription plans; enterprise terms vary.
  • Limitations: Rollout and connector eligibility can differ across accounts.

Sources & Further Reading
Claude product overview
Claude pricing

Buyer question — our analysis: Can you separate read access from write actions for each connector and revoke both centrally?

Back to Index routes

Gemini Deep Research

Gemini Deep ResearchResearch agentEvidence: Level 2 — Publicly available

Google · Assistants and research
Also: Research agent; connected-source analysis

Google documents a research workflow that proposes a plan, consults selected sources and produces a report. This entry covers Deep Research, not every Gemini capability.

Availability: Generally available to eligible accounts; access and limits vary.
Evidence: Level 2 — Publicly available
Primary users: Researchers; analysts; knowledge workers.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Chat request, editable research plan and asynchronous report generation.
  • Models: Gemini; exact model depends on the offered experience.
  • Tools and integrations: Google Search; Gmail and Drive when the Workspace connection is enabled.
  • Browser / computer use: Web research; arbitrary browser transactions or desktop control not established.
  • API / agent building / multi-agent: This consumer workflow’s API and multi-agent controls are not clearly documented.
  • Memory / persistence: Research conversations and reports; durable action memory not established.
  • Approvals and permissions: Review the research plan; connected Workspace permissions govern source access.
  • Auditability: Report sources; a complete machine-readable action log is not established.
  • Deployment: Google-hosted Gemini experience.
  • Pricing: Account/plan limits apply; verify current entitlement.
  • Limitations: Source retrieval and synthesis do not guarantee completeness or factual accuracy.

Sources & Further Reading
Use Deep Research in Gemini Apps

Buyer question — our analysis: Does the report support its key conclusions with sources you can inspect and independently verify?

Back to Index routes

Microsoft 365 Copilot

Microsoft 365 CopilotWork assistantEvidence: Level 2 — Publicly available

Microsoft · Assistants and research
Also: Work assistant; enterprise knowledge

Microsoft documents an assistant grounded in organizational information and Microsoft 365 applications. Copilot Chat access does not automatically include every licensed work-data capability.

Availability: Generally available; licensed features and geographic eligibility vary.
Evidence: Level 2 — Publicly available
Primary users: Enterprise and SMB knowledge workers; Microsoft 365 administrators.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Chat and assistance inside supported productivity experiences.
  • Models: Microsoft-managed model stack; feature-specific selection varies.
  • Tools and integrations: Microsoft 365 applications, Microsoft Graph and work-data grounding.
  • Browser / computer use: Web grounding is documented; general desktop action scope is not established here.
  • API / agent building / multi-agent: Custom agent development belongs in the separate Copilot Studio entry.
  • Memory / persistence: Work context; durable task-memory guarantees not clearly documented.
  • Approvals and permissions: Existing organizational access and administrator policies apply.
  • Auditability: Verify feature-specific audit coverage and retention with the administrator.
  • Deployment: Microsoft cloud and Microsoft 365 clients.
  • Pricing: Subscription licensing; distinguish Copilot Chat from paid Copilot entitlements.
  • Limitations: Licensing, information permissions and tenant configuration shape the available experience.

Sources & Further Reading
Microsoft 365 Copilot overview

Buyer question — our analysis: Are existing document permissions appropriate before you expand access to organizational information?

Back to Index routes

Computer and browser use

These tools operate through interfaces. The execution environment, authenticated session and transaction permissions are as important as the model.

OpenAI computer use

OpenAI computer useComputer-use APIEvidence: Level 2 — Publicly available

OpenAI · Computer and browser use
Also: Computer-use API; browser automation building block

OpenAI documents model-directed interaction with computer interfaces. Developers provide the execution environment and execute the requested actions; this is not a complete deployed worker.

Availability: Unclear release label; public API documentation specifies supported configurations.
Evidence: Level 2 — Publicly available
Primary users: Developers; automation and security teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: API action loop using screenshots, mouse/keyboard operations or code-driven control.
  • Models: Supported OpenAI models and tool variants; check current compatibility.
  • Tools and integrations: Developer-provided browser or desktop environment.
  • Browser / computer use: Both are documented; action execution remains the integrator’s responsibility.
  • API / agent building / multi-agent: API building block; orchestration must be implemented separately.
  • Memory / persistence: Maintain environment state across calls; durable memory is application-owned.
  • Approvals and permissions: Use isolation, allowlists and confirmation for consequential actions.
  • Auditability: Capture actions and outcomes in the integrating application.
  • Deployment: Hosted inference with an integrator-managed execution environment.
  • Pricing: Usage-based API; runtime costs are additional.
  • Limitations: Untrusted screen content can influence actions; enforce time, cost and access boundaries.

Sources & Further Reading
OpenAI computer use guide

Buyer question — our analysis: Can your runtime prevent an action even when the model asks for it?

Back to Index routes

Claude computer use

Claude computer useComputer-use APIEvidence: Level 2 — Publicly available

Anthropic · Computer and browser use
Also: Computer-use API; UI automation building block

Anthropic documents computer-use tools that let Claude request interface actions from screenshots. The application supplies the computer environment and executes the tool requests.

Availability: Generally available toolset on Claude API/Google Cloud; earlier variants and other platforms can be beta.
Evidence: Level 2 — Publicly available
Primary users: Developers; automation and security teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: API loop: observe a screen, request actions and receive results.
  • Models: Tool-version and provider compatibility determine supported Claude models.
  • Tools and integrations: Integrator-provided computer tools and execution environment.
  • Browser / computer use: UI interaction within the supplied environment; not blanket access to a user’s machine.
  • API / agent building / multi-agent: API building block; workflow and multi-agent architecture are separate.
  • Memory / persistence: Conversation and environment state; long-term memory is application-owned.
  • Approvals and permissions: Integrator must define access boundaries and consequential-action review.
  • Auditability: Tool requests/results can be recorded by the application.
  • Deployment: Hosted model plus integrator-managed computer environment.
  • Pricing: API usage and environment costs; provider terms differ.
  • Limitations: Tool versions, platform support and latency vary; UI access introduces security risk.

Sources & Further Reading
Claude computer use documentation

Buyer question — our analysis: Which tool version is supported on your chosen provider, and who approves irreversible actions?

Back to Index routes

Browser Use

Browser UseBrowser agentEvidence: Level 2 — Publicly available

Browser Use · Computer and browser use
Also: Browser agent; browser infrastructure

Browser Use documents a Python browser-agent library and two cloud routes: a managed Agent API and browser infrastructure for an agent you operate yourself.

Availability: Generally available open-source package and cloud APIs; service-specific eligibility varies.
Evidence: Level 2 — Publicly available
Primary users: Developers; automation teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Task-driven browser interaction; API or Python integration.
  • Models: Browser Use, Google, OpenAI and Anthropic model integrations are documented.
  • Tools and integrations: Chromium; cloud browsers can connect through Playwright/Puppeteer CDP.
  • Browser / computer use: Browser automation; general desktop control is not established.
  • API / agent building / multi-agent: Agent API or bring your own agent; multi-agent guarantees not established.
  • Memory / persistence: Browser/session behavior depends on configuration; durable memory not clearly documented.
  • Approvals and permissions: Credential and browser-session scope require integrator design.
  • Auditability: Complete retention/export guarantees not clearly documented in reviewed quickstarts.
  • Deployment: Local library or hosted browser/agent services.
  • Pricing: Open-source route has model/runtime costs; cloud tariff not established here.
  • Limitations: Authenticated browser sessions can expose sensitive data and write actions.

Sources & Further Reading
Browser Use cloud quickstart
Browser Use open-source quickstart

Buyer question — our analysis: Can you restrict domains, credentials and transactions before granting a browser session?

Back to Index routes

Coding agents

Compare the agent’s execution environment and review policy. Producing a patch is different from testing it, approving it and safely merging it.

GitHub Copilot

GitHub CopilotCoding agentEvidence: Level 2 — Publicly available

GitHub / Microsoft · Coding agents
Also: Coding agent; CLI; repository workflows

GitHub documents Copilot agents that can work on code in cloud, CLI and application environments. Cloud tasks can create branches and pull requests for review.

Availability: Generally available product; agent features, plans and policies vary.
Evidence: Level 2 — Publicly available
Primary users: Software developers; engineering teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Interactive or asynchronous repository tasks; scheduled automation where enabled.
  • Models: Available models depend on plan and agent environment.
  • Tools and integrations: Repository files, commands, GitHub workflows and MCP.
  • Browser / computer use: Computer use exists in supported CLI/app configurations; do not infer universal access.
  • API / agent building / multi-agent: SDK uses the CLI; supported application workflows can run parallel tasks.
  • Memory / persistence: Agent sessions and repository context; durable memory scope varies.
  • Approvals and permissions: Permission modes, repository access and configured review policies matter.
  • Auditability: Session activity and repository changes; validate retention/export requirements.
  • Deployment: Local CLI/app or GitHub cloud environments.
  • Pricing: Plan charges, AI credits and applicable Actions consumption.
  • Limitations: Generated code and review decisions require validation; configured policies can permit autonomous actions.

Sources & Further Reading
Responsible use of Copilot agents
Copilot cloud-agent automations

Buyer question — our analysis: Who can initiate a task, authorize credentials and merge its changes under your repository policy?

Back to Index routes

Codex

CodexCoding agentEvidence: Level 2 — Publicly available

OpenAI · Coding agents
Also: Coding agent; developer harness

OpenAI documents Codex as an agent for technical work across supported clients and execution environments. Its security boundaries depend on sandbox and approval configuration.

Availability: Generally available; plan and environment restrictions apply.
Evidence: Level 2 — Publicly available
Primary users: Software developers; engineering teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Delegated technical tasks; CLI, IDE and cloud experiences.
  • Models: OpenAI models; availability depends on plan and configuration.
  • Tools and integrations: Files, commands and supported app/MCP tools.
  • Browser / computer use: Environment-dependent; not automatic access to every connected computer.
  • API / agent building / multi-agent: Developer integration routes exist; exact orchestration depends on implementation.
  • Memory / persistence: Task context and environment state; long-term retention varies.
  • Approvals and permissions: Sandbox, network policy and approval policy work together; automatic review can replace some prompts.
  • Auditability: Tool/approval activity; confirm client-specific export and retention.
  • Deployment: Local CLI/IDE and OpenAI cloud environments.
  • Pricing: ChatGPT plan usage and extra credits; API-key use billed separately.
  • Limitations: Approval does not guarantee safety; cloud and local credential boundaries differ.

Sources & Further Reading
Codex with your ChatGPT plan
Agent approvals and security

Buyer question — our analysis: Does the effective sandbox match the permissions you believe you granted?

Back to Index routes

Claude Code

Claude CodeCoding agentEvidence: Level 2 — Publicly available

Anthropic · Coding agents
Also: Coding agent; command-line workflows

Anthropic documents Claude Code as an agent that reads repositories, edits files and runs commands. It is available through terminal, IDE, desktop and browser experiences.

Availability: Generally available; platform/provider features vary.
Evidence: Level 2 — Publicly available
Primary users: Software developers; engineering teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Interactive and delegated development tasks.
  • Models: Claude; provider and account configuration determine access.
  • Tools and integrations: Repository and command tools; integrations vary by environment.
  • Browser / computer use: Browser client does not itself prove arbitrary website or desktop control.
  • API / agent building / multi-agent: Integration and orchestration scope varies; not established by these overview sources.
  • Memory / persistence: Project context; exact durable-memory guarantees not assessed.
  • Approvals and permissions: Fine-grained allow/deny rules, modes and managed policies.
  • Auditability: Permission configuration can be version-controlled; log-export coverage needs verification.
  • Deployment: Local development clients and hosted experiences.
  • Pricing: Claude subscription or usage-billed provider access, depending on setup.
  • Limitations: Auto mode uses a classifier; it is not a guarantee of human approval for every action.

Sources & Further Reading
Claude Code overview
Claude Code permissions

Buyer question — our analysis: Which commands and file paths are allowed, and are managed policies enforced across every client?

Back to Index routes

Amazon Q Developer

Amazon Q DeveloperAWS development assistantEvidence: Level 2 — Publicly available

Amazon Web Services · Coding agents
Also: AWS development assistant; code modernization

AWS documents Q Developer for application development and AWS assistance. Its documentation now directs IDE users toward Kiro ahead of the announced plugin support end date.

Availability: Generally available; Eclipse integration is preview. IDE support ends April 30, 2027.
Evidence: Level 2 — Publicly available
Primary users: Developers; AWS operations teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Chat, code generation, scanning and development workflows.
  • Models: Built on Amazon Bedrock; pricing page identifies Claude model access.
  • Tools and integrations: Supported IDEs, AWS console/websites and chat applications.
  • Browser / computer use: Console/chat access; general browser or desktop control not established.
  • API / agent building / multi-agent: Not a general agent-building API in these sources.
  • Memory / persistence: Durable task memory not clearly documented.
  • Approvals and permissions: AWS IAM permissions; Teams/Slack channel guardrails; context-specific access.
  • Auditability: Pro administration features depend on the sign-in route.
  • Deployment: Hosted AWS service with client integrations.
  • Pricing: Free tier; Pro listed at US$19 per user/month, with limits and additional transformation charges.
  • Limitations: Plan migration before IDE support ends; quotas and sign-in eligibility differ.

Sources & Further Reading
What is Amazon Q Developer?
Amazon Q Developer pricing

Buyer question — our analysis: Will the planned IDE transition change your controls, workflow or licensing?

Back to Index routes

Enterprise workflow agents

These offerings connect to business processes and enterprise data. Check the identity behind each action, contracted features and escalation paths.

Microsoft Copilot Studio

Microsoft Copilot StudioAgent-development platformEvidence: Level 2 — Publicly available

Microsoft · Enterprise workflow agents
Also: Agent-development platform; enterprise workflow

Microsoft documents a platform for configuring agents, their knowledge, tools and workflows. Agent harness choice changes the balance between generative reasoning and rule-based execution.

Availability: Generally available product; selected harnesses and features have separate availability.
Evidence: Level 2 — Publicly available
Primary users: Enterprise developers; operations teams; administrators.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Conversation and configured workflow execution; triggers depend on setup.
  • Models: Harness/model options vary; not one fixed underlying model.
  • Tools and integrations: Actions, workflows and enterprise channels, including Teams and websites.
  • Browser / computer use: Connected actions; general desktop control not established here.
  • API / agent building / multi-agent: Agent building and agent handoffs are documented.
  • Memory / persistence: Agent/workflow context; durable memory scope not clearly documented.
  • Approvals and permissions: Automatic agent identities; configure authentication, connector access and governance.
  • Auditability: Verify tenant monitoring, export and retention for the selected harness.
  • Deployment: Microsoft-hosted platform and connected enterprise systems.
  • Pricing: Licensed/consumption packaging; exact current tariff not established in reviewed sources.
  • Limitations: Identity-based Conditional Access enforcement differs by channel; verify Teams versus other channels.

Sources & Further Reading
Copilot Studio agent overview
Agent identities and governance

Buyer question — our analysis: Which identity performs each action, and does it have more access than the requesting user?

Back to Index routes

Salesforce Agentforce

Salesforce AgentforceSales and customer serviceEvidence: Level 2 — Publicly available

Salesforce · Enterprise workflow agents
Also: Sales and customer service; enterprise workflows

Salesforce describes agents that use enterprise data and configured actions. Builder tools and subagents support specialized business tasks within the Salesforce ecosystem.

Availability: Generally available product; editions, add-ons and individual features vary.
Evidence: Level 2 — Publicly available
Primary users: Sales, service and operations teams; Salesforce administrators.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Employee/customer interactions and configured background workflows.
  • Models: Atlas Reasoning Engine; exact underlying models not established here.
  • Tools and integrations: Flow, Apex, prompts, JavaScript and MuleSoft/API connections.
  • Browser / computer use: Business-system actions; general browser/desktop control not established.
  • API / agent building / multi-agent: Builder and specialized subagents documented; implementation boundaries vary.
  • Memory / persistence: Vendor describes retained context for long-running work; confirm actual retention.
  • Approvals and permissions: Verify action authorization and escalation before deployment.
  • Auditability: Builder testing and Digital Wallet usage tracking; full action-log requirements need verification.
  • Deployment: Salesforce cloud with connected systems.
  • Pricing: Flex Credits, conversations or per-user packaging; other platform consumption may be additional.
  • Limitations: A vendor’s hypothetical cost example is not measured customer savings.

Sources & Further Reading
Agentforce product information
Agentforce pricing

Buyer question — our analysis: What is the cost per completed, verified workflow after retries and all platform charges?

Back to Index routes

ServiceNow AI Agents

ServiceNow AI AgentsEnterprise workflowEvidence: Level 1 — Announced

ServiceNow · Enterprise workflow agents
Also: Enterprise workflow; multi-agent orchestration

ServiceNow describes AI Agent Studio and an Orchestrator for teams of agents. Its product page also describes connections to external tools and agents.

Availability: Unclear at feature level; verify contracted entitlement and rollout.
Evidence: Level 1 — Announced
Primary users: Enterprise IT, service and operations teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Vendor-described conversational and workflow execution.
  • Models: Not clearly documented in the reviewed product page.
  • Tools and integrations: Flow actions, subflows, scripts, skills, MCP and A2A connections.
  • Browser / computer use: Web/desktop agent capabilities are marketed; exact supported scope is not established.
  • API / agent building / multi-agent: Studio builds agents; Orchestrator coordinates them, according to the vendor.
  • Memory / persistence: Not clearly documented.
  • Approvals and permissions: Vendor describes guardrails and governance; verify enforced action-level controls.
  • Auditability: Control Tower is marketed for governance; export/retention details not established.
  • Deployment: ServiceNow platform; verify supported hosting and data boundaries.
  • Pricing: Not publicly documented in the reviewed source; request entitlement and pricing details.
  • Limitations: This entry is supported by a vendor product description, not a verified deployment record.

Sources & Further Reading
ServiceNow AI Agents product information

Buyer question — our analysis: Which described capabilities are enabled in your contracted instance today, and can you test the controls?

Back to Index routes

Automation and orchestration

These platforms combine model-selected actions with workflow structure. Examine retries, approval gates and the boundary between deterministic and generative steps.

AI by Zapier

AI by ZapierWorkflow automationEvidence: Level 2 — Publicly available

Zapier · Automation and orchestration
Also: Workflow automation; tool-calling AI step

Zapier documents the migration of standalone Agents into AI by Zapier. Reasoning and tool calling become steps inside a broader Zap workflow.

Availability: Generally available documented migration path; account availability can vary.
Evidence: Level 2 — Publicly available
Primary users: SMB and business operations teams; workflow builders.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Triggered workflow with deterministic branches and AI-driven tool calls.
  • Models: Supported model tiers; organization policy can restrict them.
  • Tools and integrations: Configured Zapier tools and connected applications.
  • Browser / computer use: Application actions; general browser/desktop control not established.
  • API / agent building / multi-agent: Workflow/AI-step building; dedicated multi-agent orchestration not established.
  • Memory / persistence: Per-run context; durable memory equivalence to legacy Agents not assumed.
  • Approvals and permissions: Publishing/admin controls and task guardrails; runtime human approval needs separate design.
  • Auditability: Zap History exposes tool calls and task consumption.
  • Deployment: Zapier cloud with connected services.
  • Pricing: Task-based billing; do not use legacy Agents activity pricing.
  • Limitations: Migration changes architecture and billing; a publishing approval is not per-action review.

Sources & Further Reading
Migrating from Agents to AI by Zapier

Buyer question — our analysis: Which sensitive tool calls require runtime approval, rather than only approval to publish the Zap?

Back to Index routes

Make AI Agents (New)

Make AI Agents (New)Automation/orchestrationEvidence: Level 2 — Publicly available

Make · Automation and orchestration
Also: Automation/orchestration; low-code agent building

Make documents agents that select tools within its automation platform. The current New offering is in open beta, distinct from assumptions about the older experience.

Availability: Beta — open beta; functionality and pricing may change.
Evidence: Level 2 — Publicly available
Primary users: Workflow builders; SMB and operations teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Scenario-based invocation and model-selected tool execution.
  • Models: Make AI Provider; custom OpenAI, Gemini or Claude connections on paid plans.
  • Tools and integrations: Modules, scenarios, MCP tools and other agents.
  • Browser / computer use: Connected tools; general UI control not established.
  • API / agent building / multi-agent: Agent building; agents can be supplied as tools.
  • Memory / persistence: Knowledge files are documented; conversation persistence needs configuration verification.
  • Approvals and permissions: Connection scope applies; action-specific approval guarantees not established here.
  • Auditability: Run reasoning/trace information is documented.
  • Deployment: Make cloud with connected services.
  • Pricing: Make AI Provider on all plans; custom providers require a paid plan and provider costs may apply.
  • Limitations: Beta behavior can change; knowledge files should not be confused with reliable task memory.

Sources & Further Reading
Introduction to Make AI Agents (New)
Make pricing

Buyer question — our analysis: Can you trace a failed run to a specific tool call and prevent its automatic repetition?

Back to Index routes

n8n AI Agent

n8n AI AgentWorkflow automationEvidence: Level 2 — Publicly available

n8n · Automation and orchestration
Also: Workflow automation; agent-building node

n8n documents an AI Agent node that chooses connected tools to carry out a task. Human review can be attached to all tools or selected tools.

Availability: Generally available documented workflow node; hosting features depend on plan.
Evidence: Level 2 — Publicly available
Primary users: Developers; operations and automation teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Workflow execution with model-selected tool calls.
  • Models: Connected chat-model node; provider selection is configurable.
  • Tools and integrations: Connected workflow tools and external APIs.
  • Browser / computer use: Through configured tools; native general desktop control not established.
  • API / agent building / multi-agent: Workflow agent building; subagents can be connected as tools.
  • Memory / persistence: Workflow-dependent; persistent-memory configuration not assessed here.
  • Approvals and permissions: Human review pauses a selected tool before execution; approvals use configured channels.
  • Auditability: Workflow execution records; verify retention and sensitive-data handling.
  • Deployment: Cloud or self-hosted configurations.
  • Pricing: Cloud subscriptions; self-hosted packaging and license terms vary.
  • Limitations: Credentials and human-review wiring must be configured; approval prompts require suitable instructions.

Sources & Further Reading
n8n AI Agent node
Human review for tools
n8n pricing

Buyer question — our analysis: Does a denied approval reliably prevent the write action, including inside a subagent?

Back to Index routes

UiPath Agents and Maestro

UiPath Agents and MaestroEnterprise automationEvidence: Level 2 — Publicly available

UiPath · Automation and orchestration
Also: Enterprise automation; agent/RPA orchestration

UiPath documents agents working alongside deterministic automation and humans, with Maestro orchestration. Escalations can route business exceptions or approval decisions to people.

Availability: Generally available core offering; voice, memory and other capabilities include previews.
Evidence: Level 2 — Publicly available
Primary users: Enterprise automation developers; process and operations teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Agent execution inside orchestrated business processes.
  • Models: Supported models vary by configuration and data residency.
  • Tools and integrations: Activities, API workflows, automations, functions and MCP servers.
  • Browser / computer use: Through configured automation/RPA tools; not an intrinsic permission to control desktops.
  • API / agent building / multi-agent: Low-code/coded agents and orchestration of agents and workflows.
  • Memory / persistence: Agent memory is preview; escalation resolutions can be reused when configured.
  • Approvals and permissions: Action Center escalation apps; tenant governance and tool permissions need configuration.
  • Auditability: Agent traces and evaluation features are documented.
  • Deployment: Automation Cloud with connected automation environments.
  • Pricing: Platform licensing; obtain entitlement and consumption details.
  • Limitations: Escalation prompts and routing require design; agent scores are not end-to-end success rates.

Sources & Further Reading
About UiPath agents
UiPath escalations
UiPath pricing

Buyer question — our analysis: Which steps remain deterministic, and which exceptions must stop for an accountable human?

Back to Index routes

Agent frameworks and infrastructure

These are building blocks for developers and platform teams. They do not supply a finished, validated workflow or appropriate permissions automatically.

LangGraph

LangGraphAgent-development frameworkEvidence: Level 2 — Publicly available

LangChain · Agent frameworks and infrastructure
Also: Agent-development framework; durable orchestration

LangChain documents LangGraph as a low-level orchestration framework for stateful agents. It provides execution structure rather than a finished business application.

Availability: Generally available open-source framework; managed services have separate terms.
Evidence: Level 2 — Publicly available
Primary users: Developers; AI platform teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Developer-defined graphs combining deterministic and model-driven steps.
  • Models: Developer-selected models; framework is not tied to one model.
  • Tools and integrations: Application-defined tools; LangSmith tracing and deployment integrations.
  • Browser / computer use: Through application tools; not supplied as universal desktop access.
  • API / agent building / multi-agent: Programmatic agent and workflow building; architecture is developer-defined.
  • Memory / persistence: Checkpointing, durable execution and short/long-term memory patterns.
  • Approvals and permissions: Human-in-the-loop patterns; application must enforce authorization.
  • Auditability: State inspection and LangSmith tracing; configuration determines coverage.
  • Deployment: Developer-managed runtime or managed deployment options.
  • Pricing: Open-source framework; model, hosting and observability costs are separate.
  • Limitations: Persistence does not guarantee correct recovery or exactly-once external actions.

Sources & Further Reading
LangGraph overview

Buyer question — our analysis: What happens if a task resumes after a tool already performed an irreversible action?

Back to Index routes

CrewAI

CrewAIMulti-agent frameworkEvidence: Level 2 — Publicly available

CrewAI · Agent frameworks and infrastructure
Also: Multi-agent framework; workflow orchestration

CrewAI documents Crews for teams of agents and Flows for event-driven orchestration. Its managed platform extends the framework; neither label independently proves production reliability.

Availability: Generally available open-source framework; AMP is a separate managed offering.
Evidence: Level 2 — Publicly available
Primary users: Developers; AI and automation teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Role-based agent collaboration and developer-defined flows.
  • Models: Configurable model providers; verify compatibility for the chosen configuration.
  • Tools and integrations: Developer-defined tools and framework integrations.
  • Browser / computer use: Tool-dependent; no universal desktop-control guarantee.
  • API / agent building / multi-agent: Programmatic building; Crews are explicitly multi-agent.
  • Memory / persistence: Stateful Flows; durable-memory scope needs configuration review.
  • Approvals and permissions: Human-approval and access boundaries must be verified in the chosen implementation.
  • Auditability: Managed monitoring capabilities and local instrumentation differ.
  • Deployment: Developer-managed framework; managed AMP option.
  • Pricing: Open-source route plus model/runtime costs; managed-platform terms are separate.
  • Limitations: Adding agents increases coordination complexity; vendor production-readiness language is not independent evidence.

Sources & Further Reading
CrewAI introduction
CrewAI managed platform introduction

Buyer question — our analysis: Can you identify which agent caused an action and stop the workflow before it propagates?

Back to Index routes

Microsoft Agent Framework

Microsoft Agent FrameworkAgent-development frameworkEvidence: Level 2 — Publicly available

Microsoft · Agent frameworks and infrastructure
Also: Agent-development framework; orchestration

Microsoft documents Agent Framework as the successor to AutoGen and Semantic Kernel. It combines agents, tool connections and workflow orchestration; language support is not uniform.

Availability: Unclear overall release label; public packages exist, and Go support is public preview.
Evidence: Level 2 — Publicly available
Primary users: Developers; AI platform teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Developer-defined agents, harnesses and graph/functional workflows.
  • Models: Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic and Ollama integrations.
  • Tools and integrations: Function tools and MCP clients.
  • Browser / computer use: Through configured tools; no universal UI-control entitlement.
  • API / agent building / multi-agent: Programmatic agent and workflow composition.
  • Memory / persistence: Sessions and context providers; harness includes memory/context management.
  • Approvals and permissions: Tool-approval mechanisms; application authorization remains necessary.
  • Auditability: Middleware and observability integrations.
  • Deployment: Developer-managed runtime and chosen provider infrastructure.
  • Pricing: Open-source framework; provider and infrastructure costs are separate.
  • Limitations: Go preview lacks some features; do not assume examples transfer unchanged across languages.

Sources & Further Reading
Microsoft Agent Framework overview

Buyer question — our analysis: Which required features are supported in your language and version, rather than only in another SDK?

Back to Index routes

Amazon Bedrock AgentCore

Amazon Bedrock AgentCoreAgent infrastructureEvidence: Level 2 — Publicly available

Amazon Web Services · Agent frameworks and infrastructure
Also: Agent infrastructure; runtime; managed tools

AWS documents AgentCore as modular infrastructure for running and governing agents. Runtime, memory, identity, gateway and managed tools are separate capabilities, not one preconfigured worker.

Availability: Generally available documented services; components and regions have separate availability.
Evidence: Level 2 — Publicly available
Primary users: Developers; enterprise AI platform and security teams.
Last reviewed: 2026-10-08

  • Interaction and autonomy: Application/API-driven agent execution; optional managed harness.
  • Models: Bedrock and other supported providers, including OpenAI-compatible interfaces.
  • Tools and integrations: Gateway, MCP/function tools, Browser and Code Interpreter.
  • Browser / computer use: Managed Browser tool; general desktop control not established.
  • API / agent building / multi-agent: Runtime/API integration; application defines agent coordination.
  • Memory / persistence: Short- and long-term memory services; retention/configuration must be selected.
  • Approvals and permissions: Identity and isolation controls; application-owned action approvals.
  • Auditability: Observability services; coverage and retention require setup.
  • Deployment: AWS-managed runtime with isolated sessions.
  • Pricing: Usage-based by component; model and connected-service charges may be additional.
  • Limitations: Managed infrastructure does not establish application reliability or appropriate authorization.

Sources & Further Reading
What is Amazon Bedrock AgentCore?
AgentCore pricing
AgentCore release notes

Buyer question — our analysis: Can you trace identity, tool execution and cost across every service in a completed task?

Back to Index routes

Before you grant access

  1. Define the outcome: decide what a completed, verified task means before comparing demos or scores.
  2. Map identities and permissions: identify the account behind every read, write, tool call and authenticated browser session. Test allowed and blocked actions.
  3. Place approval gates: require review where money, external communications, destructive changes or sensitive data are involved. Verify that denial prevents execution.
  4. Test recovery: interrupt runs, inspect retries and confirm that resuming does not repeat an irreversible action.
  5. Calculate full cost: include model usage, platform credits, retries, infrastructure, observability and human exception handling.
  6. Check the record: establish who acted, which tool ran, what changed and how long logs and credentials persist.

Scope and maintenance

The cohort prioritizes recognizable products with a clear operational role and accessible official information. It is a starting set, not an exhaustive inventory. Assistant experiences and developer APIs are separate where their deployment and control responsibilities differ; current replacements are used instead of duplicating retired names.

Each profile follows the same fields and stable product anchor. Future reviews should check source availability, product identity, release status, pricing basis and action controls; record material changes before advancing the review date. A source check is not a fresh independent product test.

No customer counts, measured savings, benchmark rankings or undocumented commercial relationships are included. Prices are shown as billing models unless a current official figure was verified; taxes, contracts, region and add-on charges can change the total.

For interpretation and deployment coverage, explore Analysis, Infrastructure and Risk & Security. Learn about The Agentic Current or read about The Weekly Current.

Editorial disclosure: AI assisted research, drafting, and production. Reported capabilities are attributed to the sources in each profile; the buyer questions are The Agentic Current’s analysis.