Skip to main content

Why Agentic Contract Model?

The Agentic Contract Model (ACM) is a spec-first contract layer that makes AI agent systems auditable, resumable, and enterprise ready. The v0.5.0 release bundles a production grade Node.js reference implementation with typed contracts, replay bundles, and a governance surface designed for regulated industries.

What problems does ACM solve?

  • Unverifiable plans → Plans are generated as typed task graphs with context hashes, rationale, and verification guards.
  • Opaque execution → The runtime records every guard, policy decision, and tool call inside an append-only ledger that can be replayed later.
  • LLM drift → Planning is deterministic, alternatives are stored, and existing plans can be replayed without re-querying the LLM.
  • Compliance gaps → Policy and verification engines enforce safety gates before and after every task, while replay bundles ship the entire audit trail.
  • Resume failures → Checkpoints allow long-running flows to resume exactly where they stopped.

v0.5.0 highlights

  • ✅ Fully typed capability maps and registries
  • ✅ Deterministic planner with Plan-A/Plan-B alternatives and safe fallbacks
  • ✅ Resumable execution with checkpoint stores and replay bundles
  • ✅ Nucleus governance layer for LLM usage, streaming, and context directives
  • ✅ Reference experiences: five deterministic workflows and the ACM AI Coder TUI
  • ✅ MCP, LangGraph, and Microsoft Agent Framework adapters

Versioning — This documentation tracks ACM v0.5.0. Future versions will appear beside this one via Docusaurus doc versioning, so teams can adopt features at their own pace.

How the docs are organised

SectionWhat you will learn
Get StartedInstall the framework, run demos, and ship your first agent.
Core ConceptsUnderstand goals, capabilities, plans, tasks, tools, and ledgers.
PackagesDeep dives into each npm package in the monorepo.
ScenariosFive guided workflows that show ACM in action.
AI CoderOperate the full interactive developer assistant.
IntegrationsWire ACM into LangGraph, MSAF, MCP servers, and custom context providers.
GovernanceEnforce policies, verification, replay, and resumable execution.
SpecificationFollow the spec, whitepaper, and implementation plans.
ContributeBecome a maintainer, run tests, and publish packages.

ACM pillars

  1. Plan — Generate multiple plan candidates with structured tool-call envelopes and rationale.
  2. Execute — Run plans deterministically with guard evaluation, policies, verification, and streaming.
  3. Replay — Export full decision memories for compliance, analytics, and reproducibility.
  4. Integrate — Plug into existing orchestrators (LangGraph/MSAF), retrieval layers (MCP), and developer tooling (AI Coder).

Plan → Execute → Replay flow

Next steps