Durable evidence over chat memory
Important decisions, validation, review state, and follow-ups live in durable records, not private transcripts.
Opinionated SDLC
Consistent by default. Adaptable by design.
AgentFlow encodes a baseline SDLC so adopters do not need to reinvent governance for every agent, repository, or harness. It is strict where safety matters and extensible where teams need local fit.
One accountable executor
A single owner keeps routine delivery coherent even while moving through distinct responsibilities and review boundaries.
Increase intelligence per decision
Add advisory review, councils, parallel discovery, isolated spikes, routed agents, or human gates only when they improve a decision—not to maximize agent count.
Human authority where it matters
High-assurance paths keep explicit human approval for security, acceptance, and release-sensitive work.
Process does not imply permission
A workflow profile never silently grants authority to edit, publish, deploy, or perform an external action.
Truthful collaboration claims
Self-review is not independent review, and a model identifier is not evidence that a routed runtime or separate agent performed the work.
What adopters can change
Teams can extend roles, labels, workflow paths, gates, validators, release policies, skills, settings, and harness adapters.
Extensions must preserve provenance, high-assurance approval, readiness rules, no-secret evidence, and follow-ups over hidden TODOs.
What stays stable
The core protects durable evidence, explicit ownership, path-aware readiness, release assignment, action boundaries, and clear human gates. Optional extensions add capability without silently changing that contract.