darkmice

Acceptance-first agent delivery

Evidence: Case Study · Clean-room Demo · Public product surface

简体中文 · All cases

The problem

An agent producing a plausible answer is not the same as software being delivered. Real delivery needs durable task state, resumable execution, visible artifacts, independent validation, and a final decision owned by a person or policy—not by the agent that wrote the code.

System boundary

Conversation / project intent
            │
            ▼
  durable dispatch ledger ─────► dependency + timer scheduler
            │                                │
            ▼                                ▼
  local execution edge ◄──────── executor capability contract
            │
            ├────► artifacts and changed-file evidence
            └────► runtime events and failure signals
                              │
                              ▼
                 independent validation gate
                              │
                       accept / reject / retry

The control plane owns intent, dispatch state, retries, and acceptance. The edge owns the user’s workspace and adapts an available coding runtime. Validators consume outputs; they do not silently redefine the task.

Decisions and trade-offs

• Conversation as the entry point. Human instructions, agent work, artifacts, and acceptance stay in one auditable flow. This reduces context loss but requires a clear event contract.
• Persistent dispatches, not in-memory jobs. A durable ledger makes retries and reconnects explainable. The cost is more explicit state-machine work.
• Capability-based executors. Product semantics remain stable while coding runtimes vary. Provider-native events must be projected without erasing important differences.
• Acceptance as a separate gate. Completion is not inferred from a final message or a green command. Evidence must satisfy the declared checks.

Failure paths treated as product behavior

Failure	Required behavior
Connection drops mid-run	Reconnect to the same dispatch; do not duplicate the job.
Worker stops heartbeating	Expire its lease, preserve the ledger, and make retry state visible.
Agent says “done” with no artifact	Keep the delivery gate unresolved.
A validator fails	Preserve logs and failed evidence; do not overwrite history with a retry.
User changes a constraint	Record the new constraint and apply it to subsequent work.

What can be verified publicly

• The public Studio repository exposes the product-facing command-center work.
• The Evidence Lab is a clean-room, synthetic state-machine demo of dispatch, dependency, retry, intervention, evidence, and acceptance behavior.
• The production implementation, internal schemas, and operating data remain private. Claims about those details are intentionally not presented as open-source proof.

What this demonstrates

Product definition across human and agent roles, durable orchestration, protocol design, local/cloud boundary design, failure-state UX, and evidence-based acceptance.