Open-source process layer

Make AI-assisted delivery reviewable—and easy to resume.

AgentFlow adds explicit roles, durable evidence, validation gates, resumable handoffs, and human authority where risk demands it—around the repository and tools you already use.

Stable release · v1.0.0The first stable AgentFlow release, available from the tagged source checkout.

Why AgentFlow

Faster coding still needs a dependable delivery system.

When implementation accelerates, scope, architecture, validation, review, and handoff become the scarce work. AgentFlow makes those decisions inspectable without turning every task into ceremony.

Clear scopeAcceptance criteria and path selection before implementation.
Durable proofIssue, role-pass, validation, and PR evidence outlive the session.
Safer handoffsOwnership, context, boundaries, and next action remain explicit.

The working loop

Frame. Decide. Build. Verify. Review. Release.

01 Frame02 Decide03 Build04 Verify05 Review06 Release

Roles separate kinds of judgment without requiring a different person or agent for every step. Single-agent execution is the baseline; advisers and routed agents are optional and evidenced.

AgentFlow 1.0

A stable delivery system from adoption to release.

Version 1.0 brings the workflow, evidence model, validation surface, runtime adapters, optional Cockpit, and release governance into one supported source release.

Workflow

Paths, roles, and human gates

Four risk-aware workflow profiles guide nine named responsibilities while action authority and high-assurance approval remain explicit.

Evidence

Durable, portable delivery context

Requirements, decisions, validation, handoffs, lifecycle events, and PR evidence can be resumed across sessions and tools.

Assurance

Validators, evals, and outcome signals

Deterministic checks cover workflow state, releases, skills, adapters, evidence, collaboration, and executable agent evaluations.

Adoption

Review-first setup and updates

Read-only inspection, ownership-aware sync, conflict detection, and explicit approval preserve project-controlled decisions.

Integration

Harness-neutral core, native adapters

Claude Code, Codex, Agy, and Pi share one canonical SDLC definition without making generated harness surfaces authoritative.

Operations

Cockpit and release governance

Optional goal-centered views complement GitHub-backed evidence, release planning, validation, and closeout.

Product boundary

A process layer, not another development platform.

  • Does not generate your application.
  • Does not replace GitHub, your agents, or your technology stack.
  • Does not require many agents for routine work.
  • Does keep consequential approvals with humans.
  • Does turn hidden workflow assumptions into inspectable contracts.