concepts
Mission Sizing
A candidate must satisfy at least 3 of 5 criteria to qualify as a Mission:
| Criterion | What It Means | Litmus Test |
|---|---|---|
| Multi-agent | Requires 2+ distinct profiles/tools | Would a single agent choke on this? |
| Multi-session | Expected to span multiple chat sessions or Kanban lifecycles | Would it survive a human going to sleep and returning? |
| Multi-objective | Has 2+ distinct outcomes that must be achieved | Can you describe "done" in 3+ independent statements? |
| Coordination required | Outcomes have dependencies, sequencing, or parallel paths | Does order matter? Can some work run in parallel? |
| Method ambiguous | The "how" is not obvious; mission command is needed | Would a competent builder need significant judgment calls? |
Examples of real missions
Hardening authentication end to end, migrating frameworks, security audits, building a public site, implementing an observability stack.
Examples of tasks
Rewriting a README, fixing an auth bug, reviewing a PR, researching best practices, updating dependencies.
The test: If you gave this to one agent with one /goal, would they succeed? If yes, it is a task. If no, because they lack expertise, add a reviewer. Still a task. If no, because it requires coordination across systems, sessions, or tools, it is a Mission.
Sizing is advisory, never enforced. The operator decides; the system informs the decision and respects it.
Structural Limits
The architecture makes five things structurally possible and four things structurally impossible.
Possible
- Intent-driven automation: human writes intent, system decomposes.
- Persistent multi-session execution: the board survives restarts, crashes, and session boundaries.
- Observable agency costs: every card generates rows in task_runs; the After-Action Review is a view of the database.
- Composability without lock-in: any tool that speaks
/goalcan join. - Human-in-the-loop governance: the “blocked” state is residual control made operational.
Impossible
- Opaque black boxes: decomposition is visible before execution.
- Unbounded delegation: budgets gate every card.
- Single-tool lock-in: the Flight Plan compiles to universal primitives.
- Fire-and-forget automation: the system expects supervision, and the Flight Plan makes that expectation explicit.