Poesis
Open · Apache-2.0
The SAFe Agentic Framework.
The open BUILD layer: SAFe-shaped multi-agent orchestration that turns governed GSM definitions into runnable code — closing THINK → BUILD → RUN.
Why it matters
The generative bridge — and the harness that makes it trustworthy.
Definitions that generate systems.
The framework is the BUILD layer that turns governed GSM definitions into running systems — making GSM's promise, “definitions that generate systems”, literally true, and closing THINK → BUILD → RUN as one loop.
A method, run by agents, as a harness.
AI runs on intuition — exactly like we do — and methods are how intuition becomes reliable. SAFe, run by agents with a human ★ gate at every layer, is the cognitive harness that makes agentic delivery trustworthy. You don't wait for a smarter model; you make it implement a method.
What it does
Four functional components, one governed delivery loop.
Portfolio → program → iteration
Three orchestrators, one per SAFe layer, run an org chart of role agents — not a flat swarm — with a human ★ gate at every layer.
The specialist bench
A bench of specialist agents — architecture, security, QA, UX and more — dispatched into each ceremony as stateless subagents.
The deterministic harness
Step graphs validated by a deterministic harness — the gates, conditions, and ledger that keep probabilistic agents on rails.
Local-first & sovereign
The agent loop runs on your machine against your working tree; work state lives in your own files and git history, not a vendor's cloud.
Orchestration
Portfolio → program → iteration.
Three orchestrators — one per SAFe layer — run the show from a single entry point: the portfolio layer tests the strategic bet, the program layer tests the feature shape and architecture, the iteration layer tests the code that ships. It is an org chart with human authority, not a flat swarm, with a human ★ gate at every layer.
Specialist bench
A team, not a clever individual.
Each ceremony dispatches the specialists it needs — product, architecture, security, QA, UX, DevOps — as stateless subagents that read committed artifacts and return. Division of labor makes every agent small, bounded, and accurate for its role, which is what makes a pipeline of focused agents outperform a single sprawling one.
Harness
A deterministic harness keeps agents on rails.
The workflow step graphs are validated by a deterministic harness — gates, conditions, and an append-only ledger — so the probabilistic agents move through evidence, checks, and sequence until the output is trustworthy. More governed layers across the end-to-end path produce better outcomes than a method that checks only one slice.
Local-first
Local-first & sovereign.
The agent loop runs on each user's machine, in the IDE, against the real working tree — your sensitive information never leaves; only the artifacts you commit sync out. Work state lives in your own files and git history (the event log), not a vendor's cloud. Nothing is trapped in an unmanaged session.