Handbook

Platform Charter

Forge Platform exists to make agentic software delivery governable, visible, and evidence-producing across multiple products and agent workcells.

Updated

Mission

Forge Platform exists to make agentic software delivery governable, visible, and evidence-producing across multiple products and agent workcells.

The platform's core loop is:

Platform core delivery loop

How Forge governs agentic delivery from intent through bounded execution to reviewable evidence and human decision.

  1. intentThe human or team states what they want delivered or changed.
  2. Blueprint-shaped contextBlueprints structure the work with governed, reusable practice knowledge.
  3. LCDL-governed reasoningLCDL runs schema-aware planning or review tasks with predictable outcomes.
  4. human approvalA human approves the bounded plan before execution proceeds.
  5. Fleet-controlled executionFleet launches containerized jobs inside token-protected execution boundaries.
  6. agent workcell outputBounded agent workcells perform scoped work inside the operating envelope.
  7. EvidencePacketThe run emits structured evidence of what was done and produced.
  8. Lenses reviewLenses surfaces the run state, artifacts, and evidence for inspection.
  9. human decisionThe human approves, rejects, defers, or creates follow-on work.

North-star promise

From intent to evidence: Lenses shows the work, Blueprints structure it, LCDL verifies it, Fleet runs it, agents perform scoped work, and humans approve it.

Architectural stance

Forge is not an agent. Forge is the operating envelope that makes many agents governable.

Agents such as Hermes, Factory Droid, OpenClaw, Codex, Claude Code, and future tools should be modeled as bounded workcells.

Forge Campaigns are multi-repo automation efforts under one ForgeRun, driven by a campaign manifest and one or more Forge Agents (AgentRuns). Campaigns produce evidence and a human decision in Lenses — they are not separate from the run spine. See Forge Campaign and the workcells hub.

MVP objective

For one local workspace and one repo, a human can:

  1. create or receive intent,
  2. inspect the relevant Blueprint context,
  3. run governed LCDL planning or review,
  4. approve a bounded Fleet template job,
  5. optionally invoke an agent workcell,
  6. review an EvidencePacket in Lenses,
  7. decide whether to approve, reject, defer, or create follow-on work.

Guardrail

A feature belongs in the platform only if it improves a visible ForgeRun from intent toward evidence.