Saphan Protocol — the written discipline behind the engine.
A versioned, human-readable protocol for how AI agent fleets deliver work: what an order must contain, what evidence must accompany a result, and where a human decision is required before anything advances. The engine automates it. Anyone can read it.
What the protocol defines
Orders
Self-contained, paste-ready briefs. An executor receives everything it needs — scope, constraints, proof standard — and nothing it must guess. Ambiguity in the order is the dispatcher's defect, not the agent's.
Gates
Named STOP points where work halts until a human decides. Design is gated before implementation; results are gated before merge. Every gate decision records the actor. The engine can refuse; only a human can approve.
Evidence standard
A claim without proof does not pass. Builds, tests, diffs, and reproduction steps travel with the result — and a number without a definition is not evidence. Approval without evidence is unrepresentable in the system, by construction.
Decision records
Decisions outlive conversations. Architecture choices, scope rulings, and naming law live in versioned records with their reasons — so the answer to "why is it like this" is a document, not a memory.
The findings loop
Every failure becomes a rule in the same cycle it was found. The protocol is not a manifesto written once — it is case law, hardened daily against our own fleets.
Principles, verbatim
Verify before assert.
Chat is not the record.
Observe before destroy.
Push is publication, not merge.
The engine never merges.
Why it lives outside the tool
A discipline locked inside a product dies with the product. Saphan Protocol is a document first: your team can read it, adopt it, and audit against it — with or without our engine. The engine's job is to make following it effortless and breaking it visible.
Its records are designed to map onto the change-management controls organizations already answer for — the questions an ISO 27001 or SOC 2 auditor asks about who changed what, on whose approval, with what evidence.
And it is executor-agnostic by construction: agents sit behind a boundary contract. Swap the agent stack; the protocol — and the record — stays.
Status
The protocol runs in production today, governing the fleets that build Saphan itself and our other systems.
Saphan Protocol v1.0 publishes openly with the release, together with the engine and the audit pack.
Design partner inquiries
Talk to the founder — hello@saphan.ai