From requirement to verified change
One governed pipeline, from a spec to a sealed, audited change.
This is AI change management as a single continuous sequence. A requirement is scoped on the graph, graded for risk, planned, approved by a human, executed by governed agents, and sealed into an evidence ledger. Months become hours by removing the work that never needed a human, discovery, scoping, and plumbing, and by making the risky parts governed rather than skipped.
Seven stages carry the change all the way to production.
Each stage feeds the next, and each transition is sealed as it happens. The order is the point: understanding and governance come before generation, so an autonomous change is scoped correctly the first time.
- 01
Comprehensive spec intake
A requirement, PRD, or ticket becomes a complete, structured specification.
The change starts as a full intake, not a one-line prompt: the intent, the acceptance criteria, and the constraints, captured so every later stage has something precise to work from.
- 02
Impact scoped on the graph
The spec is traversed across the Engineering Memory Graph to its real blast radius.
Impacted services, APIs, events, database tables, and owners are resolved deterministically from evidence, before anyone writes code. Same graph plus same spec gives the same impact set, and every impacted-node claim links the evidence that justifies it.
- 03
Risk assessment
The scoped impact is graded for risk against graph-aware policy.
Coupling, ownership, and blast radius drive a risk read: a change that touches payment_transactions is not treated like a copy tweak. The risk grade decides which approvals and which tests the plan must include.
- 04
Test, approval, deployment, and rollback plans drafted
A governed change plan is drafted from the impact and the risk grade.
Findry drafts the test plan (unit, integration, contract, event, rollback), the approvals the policy requires, the deployment and canary strategy, and the rollback path, all scoped to what the change actually touches.
- 05
Human approval gate
A person reviews the impact, the risk, and the plan, and approves or rejects.
Nothing autonomous ships without this gate. The reviewer sees the evidence-backed impact set and the drafted plan, and the decision itself is recorded. Approval is what starts governed execution.
- 06
Governed execution
Agents dispatch, changes are reviewed, scanner gates run, and rollout is staged.
Governed AI coding agents write the change, it goes through review, policy and scanner gates evaluate in-line, and a canary rolls out through your continuous delivery, advancing on healthy signals and rolling back on breach.
- 07
Sealed into the evidence ledger
Every stage transition is sealed into an append-only, hash-chained trail.
Audit is not a report assembled afterward. Each transition, intake, impact, risk, plan, approval, dispatch, gates, rollout, is sealed as it happens into exportable, immutable evidence you can hand to a regulator.
Requirement → Impact → Risk → Plan → Approval → Execution → Sealed
The generation is deterministic where it can be: the language model writes the narrative on top of an impact set the graph already resolved from evidence. The human gate governs the leap from plan to production, and the ledger makes every step provable after the fact.
The mechanism that makes stage two deterministic is the Engineering Memory Graph. The full product surface is on the platform page.
Questions we get asked.
- How does a change start?
- A change starts as a comprehensive, structured specification, not a one-line prompt. The requirement, PRD, or ticket becomes a full intake capturing the intent, the acceptance criteria, and the constraints, so every later stage has something precise to work from and the spec can be traversed across the graph to its real blast radius.
- Who approves before agents execute?
- A human owner approves at the approval gate before anything autonomous ships. The reviewer sees the evidence-backed impact set, the risk grade, and the drafted plan, then approves or rejects, and the decision itself is recorded. Approval is what starts governed execution: nothing autonomous ships without passing this gate.
- What happens if a rollout goes wrong?
- A gated canary rolls out through your continuous delivery, advancing on healthy signals and rolling back on breach. The rollout strategy and the rollback path are drafted in advance and scoped to what the change touches, and every transition, including the rollback, is sealed as it happens into the append-only, hash-chained evidence trail.