Skip to content
Agent Forge Docs

Coding agents do not run inside a platform—how do you govern them?

Many coding agents run on a developer machine or in a separate process—they are not hosted inside a “platform runtime.” If governance is an SDK library linked into that process, it only constrains that link; start another process beside it and control is gone. To govern side effects, the gate must live on an out-of-process service surface.

Constraints

Section titled “Constraints”
flowchart LR
  Agent[Executing process] -->|action metadata| Gov[Governance service]
  Gov -->|allow/deny/ask| Agent
  Agent -->|only after allow| Side[External side effects]
  • Report tool name and parameter summary—do not execute first and bookkeep later
  • Without allow, do not execute; unmatched rules may default to deny
  • Ask enters human approval; it does not auto-allow on timeout

How Agent Forge connects

Section titled “How Agent Forge connects”

The console and runtime already speak HTTP/SSE. Cross-process scenarios reuse the same governance semantics: blocking pre-check via /govern, then HITL and audit. Architecture argument: why governance must be out-of-process.

Cross-process wiring ships as a reference demo path; whether it covers your executor depends on the integration.

Related reading: