Agents decide what should happen.Converge makes reality match it.
An AI agent can reason correctly and still leave external systems inconsistent. Converge persists the desired outcome, observes authoritative state, and deterministically repairs only the remaining safe drift.
The refund is already true in Stripe, so it is observed—not repeated.
The model can be right and the system can still end up wrong.
A refund succeeds in Stripe. The process stops before the support case updates. Retrying the whole agent risks repeating a financial side effect.
Reasoning chooses intent. Deterministic systems own execution truth.
Persist first
The desired outcome is durable before any external mutation begins.
Observe reality
Recovery starts by re-reading Stripe and Supabase, not replaying a trace.
Repair the gap
Only unsatisfied requirements are eligible for mutation.
Stop when true
A later worker sees the satisfied outcome and performs zero new mutations.
The model is not the transaction coordinator.
Gemini gets read-only tools and declares the desired business outcome. Converge owns the durable state machine, ownership checks, observation, idempotency, and repair.
Agent reasoning
Interpret request, inspect facts, choose the desired target state.
Outcome contract
Persist requirements before mutation so another worker can continue safely.
Reconciler
Observe authoritative systems and repair only unsatisfied safe drift.
Converge verifies the declared outcome. It does not replace business judgment.
The agent selects the semantic target. Converge verifies resource ownership and safely reconciles that declared target against authoritative state. Those are different guarantees, and the demo keeps them separate.
Break it on purpose.
The demo issues a real Stripe test-mode refund, intentionally stops mid-flight, then proves that Converge can recover without duplicating the side effect.
Run demo