Skip to content

slides.md

Latest commit

 

History

History
147 lines (101 loc) · 2.39 KB

slides.md

File metadata and controls

147 lines (101 loc) · 2.39 KB
Raw
title AI Agent Governance Workshop Slides
last_reviewed 2026-07-12
owner docs-team
marp true
theme default
paginate true

Introduction to AI Agent Governance

Native policy, host controls, identity, and trust

Why governance

  • Agent actions can create external side effects.
  • Prompt instructions are not an authorization boundary.
  • Policy, identity, isolation, audit, and reliability solve different problems.

Native ACS policy

An ACS manifest declares:

  • policy engines and bundles
  • intervention-point bindings
  • tool catalogs
  • transforms and evidence
  • approval
  • limits

Runtime model

from agent_control_specification import AgentControl, HostSession

runtime = AgentControl.from_path("policies/manifest.yaml")
session = HostSession(
    runtime,
    agent_id="agent-1",
    session_id="session-1",
)

Tool mediation

evaluation = session.pre_tool_call(
    tool_name="delete_file",
    args={"path": "report.txt"},
)

No side effect occurs before the pre-intervention evaluation.

Result contract

PolicyEvaluation carries:

  • verdict and reason code
  • transform and evidence
  • input and enforced identities
  • approval metadata
  • restricted audit payload

Lab 1

Run labs/lab1_first_policy.py.

  1. Inspect the manifest binding.
  2. Inspect the Rego policy.
  3. Run the scenarios.
  4. Add one deny condition.

Host controls

Policy evaluation does not replace:

  • sandbox resource and network controls
  • framework lifecycle ordering
  • circuit breakers and SLOs
  • identity and trust

AgentMesh

AgentMesh owns:

  • cryptographic identity
  • trust and reputation
  • capability grants
  • cross-agent transport

Lab 2

Complete labs/lab2_multi_agent_trust.py.

  1. Create identities.
  2. Attempt a trust handshake.
  3. Build trust.
  4. Revoke and verify denial.

Fail closed

  • Unknown tools can be denied by manifest contract.
  • Runtime and approval errors deny.
  • Action-identity mismatch denies.
  • Public errors remain sanitized.

Production checklist

  • Lint and replay manifests.
  • Test every mediated framework path.
  • Test sandbox isolation.
  • Test approval and runtime failures.
  • Restrict audit access.
  • Monitor SLOs and incident signals.

Next steps

  • docs/quickstart.md
  • docs/tutorials/03-framework-integrations.md
  • docs/tutorials/55-agent-control-specification.md
  • docs/workshop/lab-guide.md