Introduction

Agent Forge is an agent runtime with an open-source MIT kernel and out-of-process governance. These docs describe current behavior: how a request moves through the kernel, at which layer governance closes, what reliability covers, and why the design is shaped this way.

The kernel derives from an MIT-licensed upstream agent runtime; the upstream repository, locked commit, and license boundary are recorded in Baseline traceability. Other sections describe current Agent Forge behavior only and do not compare against upstream.

They are written for people who need to understand how Agent Forge works, and focus on flow, boundaries, and verification criteria. Code appears only as support; content is not organized as a source walkthrough. Except for the baseline topic, the text does not name the upstream product.

Columns

Column	What it covers	Typical use
Reliability	Pinned baseline version and hash checks, module test coverage, audit-chain and HITL metric definitions, deployment and isolation acceptance	Check version traceability and whether isolation holds
Architecture	Layering, out-of-process governance, multi-tenant RLS, the Open-Core seam, trigger chain, Effect conventions	Build the system map first, then drill into topics
Kernel	How user input becomes a tool call: message model, runLoop, Processor and LLM adaptation, tool system, config and permission overlays	Debug the main path or change kernel behavior
Governance	HITL, audit chain, breaker, RBAC, identity binding, file sandbox, retention, observation	Wire in governance or run a compliance acceptance
Why Agent Forge	Five premises usually skipped, and the mechanisms that answer them	Review the reasoning behind design trade-offs

Suggested reading order

For a technical evaluation; the system map can come earlier:

1. Reliability — confirm the pinned version and what is covered.
2. Technical architecture overview — topology, five layers, request path, governance attachment points.
3. Kernel — the main path and tool execution.
4. Governance — how each governance capability closes.
5. Architecture — layering detail and isolation boundaries.
6. Why Agent Forge — trade-offs and where they apply.

Conventions and boundaries

• The text describes current Agent Forge behavior only. Where the docs and the implementation disagree, trust agent-forge-runtime and the reproduction result.
• Chinese and English share slugs: /docs/<slug>/ and /en/docs/<slug>/. Pages without a translation fall back to the Chinese text.
• Configuration uses the AGENTFORGE_* prefix; request headers are X-AgentForge-Tenant-Id, X-AgentForge-Owner-Id, and X-AgentForge-Role.