Comparison · Checked November 6, 2026

Sentience vs. Microsoft Agent Governance Toolkit

Both start free and open source. The difference is what comes next. Sentience is one product that grows with you: governance that installs beside your agents, then enforcement and isolated execution, configured for your workflows by our team. Microsoft’s toolkit is a public-preview library your developers embed and maintain.

The differences that matter

Where Sentience is different.

  1. 01

    Who deploys it

    Sentience

    • Governor: a developer or operator, locally, beside the agent.
    • Enforcement: configured for your workflows and integrated with your existing infrastructure, on AWS, Azure, GCP or a private environment.
    • Secure Runtime: built for production from day one, custom-built on Sentience’s isolated environments; your systems connect to it.

    AGT says

    A library developers embed in their agents (Python, npm, NuGet, crates.io and Go), with deployment guides for Azure, AWS and Google Cloud.

  2. 02

    What happens on a violation

    Sentience

    • Governor: evaluates each tool call its integration sees at the execution boundary against the declared scope, then records and flags it. Observe-only and fail-open: the action proceeds.
    • Enforcement: redirects, pauses or stops the action at control points integrated with your environment.
    • Secure Runtime: integrated enforcement, scoped credentials and policy-governed egress.

    AGT says

    AllowDenyRequire approvalTransformKill switch
  3. 03

    Can it correct the agent, not just stop it?

    Sentience

    Yes, and the correction is yours. Runtime Enforcement follows your policy book, configured for each workflow: redirect an email to a draft, hand a payment to a person for approval, steer the agent back to its declared task.

    Course-correctEscalateStop

    AGT says

    Returns allow, deny or transform. A transform replaces part of the tool call’s arguments before it runs; Microsoft’s own example redacts content from an email.

  4. 04

    Where the record lives

    Sentience

    A local record of what was checked and what happened, stored as files you own. Each record names the governance profile that governed the run, by its content hash. No telemetry on the default path; governance works with the network off. Every layer includes Governor.

    AGT says

    “Tamper-evident records of every decision.”

  5. 05

    Mapping to frameworks

    Sentience

    A custom service. Your policy book can be mapped to ISO/IEC 42001 and EU AI Act requirements for your workflows. Because the map starts from your policy book, it reflects your own controls, not a generic checklist.

    ISO/IEC 42001EU AI Act

    Mapping supports your assessment. It is not a certification.

    AGT says

    OWASP Agentic Top 10NIST AI RMFEU AI ActSOC 2ISO/IEC 42001CSA ATFAARM Extended
  6. 06

    Works with

    Sentience

    Claude CodeCodexPydantic AIMCP clientsLangChainLangGraphLLM gateways

    Governor installs beside the coding agents and frameworks you already run. Nothing to migrate.

    Records stream to standard output, files or a local endpoint, so you can forward them to the guardrails, evaluation and observability tools you already use.

    Enforcement runs where your agents already run, on AWS, Azure, GCP or a private environment, alongside your models and LLM gateways.

    AGT says

    Claude Code, Codex CLI and GitHub Copilot CLI; adapters for LangChain, LangGraph, CrewAI, OpenAI Agents, PydanticAI, LlamaIndex, Haystack, Google ADK and other frameworks; MCP.

  7. 07

    What it costs to find out

    Sentience

    Nothing to start.

    Governor is free and Apache-2.0. Enforcement is configured for your workflows and priced per governed workflow; the per-workflow price drops as you scale. Secure Runtime is custom-built for your environment, starting with a proof of concept.

    pipx install sentience-governor

    AGT’s pages

    Free and open source under the MIT license. Public preview.

Microsoft Agent Governance Toolkit entries are the vendor’s own claims, from its website as of November 6, 2026. They have not been independently tested. This page makes no claim about whether Sentience and AGT can be used together.

Same call, two approaches

An invoice agent tries to email a customer.

The agent is declared to reconcile invoices. Midway through, it calls a tool that sends an email, outside its declared scope.

SentienceIllustrative: a configured policy

Seen, corrected, carries on.

  1. 1
    Governor sees it

    Captured before dispatch, flagged as outside the declared scope, recorded locally.

  2. 2
    Enforcement, configured for this workflow, corrects it

    Your policy book turns the email into a draft for review.

  3. 3
    The agent carries on

    Back to reconciling invoices. Nothing was sent.

Microsoft Agent Governance ToolkitAGT says

Advertised options.

AllowDenyRequire approvalTransform

Ask before adopting: Who writes and maintains the policies, and who supports it in production?

Before you decide

Questions to ask before adopting AGT.

Each one turns an advertised claim into something you can see working on your own agents.

  1. 01Who writes, tests and maintains our policies, and who supports it in production?
  2. 02What will change before general availability, and how will breaking changes reach our agents?
  3. 03After an action is stopped, does the run continue? Can the agent be corrected instead, following our own policy?
  4. 04How do we show later exactly which policy governed a given run?
  5. 05Who owns the integration work for each framework and coding agent we run?
  6. 06What will running it cost us in engineering time as we add workflows?

Each layer does one job

One product, every layer accounted for.

Hold any vendor, including AGT, to the same clarity.

  1. 01

    Governor does not stop or alter an action. It records and flags.

    Runtime Enforcement redirects, pauses or stops it.

  2. 02

    Governor flags actions outside the declared scope. It cannot judge an in-scope action against an objective nobody approved.

    Your policy book, enforced at runtime.

  3. 03

    A record supports investigation. It does not prevent incidents.

    Enforcement and Secure Runtime act before the action takes effect.

  4. 04

    Governor says nothing about runs it did not observe.

    Secure Runtime runs the workflow’s actions inside its isolated, governed environment, built for production from day one.

Keep comparing

More ways to evaluate.

CompareAll comparisons Sentience compared with other companies and approaches, and by use case.ApproachesRuntime governance vs. runtime enforcement What can actually stop an AI agent action, with one worked example.Use caseAgent infrastructure governance Governance as a platform capability across the agents your teams run.

Questions

Common questions.

How is Sentience different from Microsoft Agent Governance Toolkit?

Both offer a free, open-source starting point. Sentience is one product: Sentience Governor runs locally beside your agents and records and flags, and Runtime Enforcement and Secure Runtime are configured for your workflows when you need to correct or stop actions. Microsoft's Agent Governance Toolkit is an MIT-licensed library in public preview that developers embed to allow, deny or require approval for agent actions. For teams evaluating alternatives to Microsoft Agent Governance Toolkit, Sentience can be tried today, free.

What should we look for in runtime AI agent governance?

Four things: whether you can try it on your own agents before buying; whether it can correct an action, not only allow or block it; where the records live and who can read them; and whether it works with the agents you already run. Sentience Governor is free and open source, records locally in files you own, and works with Claude Code, Codex, Pydantic AI, MCP clients, LangChain and LangGraph. Runtime Enforcement adds course correction, configured for your workflows.

Is there a free way to start with agent governance?

Yes. Sentience Governor is open source under Apache-2.0 and runs locally beside your agent, with no account. It records and flags; it does not block.

Can Sentience block or correct an agent action?

Yes, with Runtime Enforcement, configured for your workflows: it redirects, pauses or stops actions at control points integrated with your environment, following your custom policy book. Sentience Governor, the free layer, is observe-only and fail-open; it records and flags.

Does Sentience map to ISO/IEC 42001 or the EU AI Act?

As a custom service, your policy book can be mapped to ISO/IEC 42001 and EU AI Act requirements. Mapping supports your own assessment; it is not a certification.

Which agents and frameworks does it work with?

Sentience Governor works with Claude Code, Codex and other coding agents, Pydantic AI, MCP clients, LangChain and LangGraph. Runtime Enforcement runs where your agents already run, on AWS, Azure, GCP or a private environment, alongside your models and LLM gateways. Records can be forwarded to the guardrails, evaluation and observability tools you already use.

Next step

See what your agents do before any demo. It’s free.

Start free with Sentience Governor

Need to stop actions, not just see them? How the layers work together