A sequence that protects the core.
Each milestone proves a difficult execution property before the next layer increases scope.
Technical validation
Prove a durable workflow, an MCP capability activity and the Tauri shell.
Validation: crash recovery and a checkpoint resumed 24 hours later.
Complete task loop
Run one research task from contract creation to accepted delivery.
Validation: a real task with execution above 50% and an acceptable delivery.
Routing and governance
Complete capability matching and all four enforcement ports.
Validation: explainable candidates, high-risk confirmations and budget compliance.
Multiple task threads
Add a lightweight queue, concurrency and recovery policy options.
Validation: state isolation and fair resource allocation under load.
Continuity and ecosystem
Introduce CRDT sync, team-aware policies and reusable execution templates.
Validation: cross-device recovery and capability reuse.
A narrow product with a complete promise.
The first release must finish the full contract-to-delivery loop before expanding collaboration or infrastructure.
Included in MVP
- Versioned task contracts and acceptance criteria
- Durable single-task execution and crash recovery
- MCP, OpenAPI and CLI capability adapters
- Routing, permissions, budgets and audit events
- Human checkpoints and delivery acceptance
- Local SQLite history and portable exports
Excluded for now
- Cloud synchronization
- Multi-user collaboration
- Capability marketplace
- Complex multi-agent negotiation
- Unconfirmed fully automatic workflows
- Organization-wide policy management
What proves the loop is working?
Feature completion is not enough. The MVP succeeds when delegation is useful, understandable and repeatable.
Task execution rate
The system completes a meaningful majority of the delegated work without continuous monitoring.
SIGNAL · 01Reviewable decisions
The user can explain every major step and intervene confidently at the right checkpoint.
SIGNAL · 02Accepted delivery
The first real task reaches an acceptable, evidence-backed result.
SIGNAL · 03Willing repetition
The user delegates another real task after accepting the first result.
SIGNAL · 04The hard parts are planned, not hidden.
The plan protects the core by using established infrastructure and explicit product boundaries.
| Risk | Control |
|---|---|
| Custom recovery complexity | Use Temporal or Restate for durable execution, replay and idempotency. |
| Tool context overload | Use enrich, retrieve, rank and restrict before the model plans. |
| Silent background execution | Expose every step and enforce explicit pause, takeover and audit paths. |
| Premature platform scope | Keep MVP local, single-device and focused on one complete task loop. |
The first proof is a user who can delegate a real task, review its evidence and accept the delivery.