Write one page of intent. A fleet of agents executes it. A debrief shows you whether your intent survived.
Every time you hand work to someone, some of what you meant gets lost. As agents take on more of the execution, that loss compounds silently across every handoff. Most agent tooling optimizes for doing more. missions.md optimizes for something different: making sure what gets done is what you meant.
The Pattern
The idea is old. Militaries call it mission command: the commander states the intent and the constraints, and trusts the unit to choose the method. NASA runs the same pattern as Mission Control: a Flight Plan, Go/No-Go checks, telemetry, and a debrief. missions.md applies that pattern to agent fleets.
You command. The system coordinates. The agents execute.
The two layers
| Layer | Owned by | What it is |
|---|---|---|
| Flight Plan | You | One page: Commander's Intent, Constraints, Success Criteria, Context |
| Hermes Kanban | The system | Cards with dependencies, budgets, verification, and human gates |
The orchestration theory lives in how the skills design the cards, not in extra artifacts you maintain.
Three Skills
| Skill | What it does | You say |
|---|---|---|
define-mission | Turns raw intent into a Flight Plan | “I want to…” |
launch-mission | Runs pre-flight checks, then compiles the Flight Plan into Kanban cards | “Launch it” |
debrief-mission | Reads the execution history and generates the After-Action Review | “Debrief” |
Skills live in the missions repo and install into ~/.hermes/skills/ for the Hermes agent.
Your First Mission
- 1.Tell your Hermes agent what you want: “I want every API route to return structured error codes, because silent auth failures are security incidents waiting to happen.”
- 2.
define-missiondrafts a Flight Plan inmissions/and asks only what it cannot infer. You approve it. - 3.Say “launch it.”
launch-missionruns the Go/No-Go checks, then creates Kanban cards assigned to your configured profiles. Dependencies gate the sequence; goal_mode judges each card against its acceptance criteria; human gates block where you said they should. - 4.Watch the Hermes dashboard, or walk away. The board survives restarts.
- 5.Say “debrief.” You get an After-Action Review: what ran, what it cost, what passed verification, and where intent drifted.
Mission or Task?
A Mission is a bounded campaign: multiple agents, multiple sessions, real coordination, judgment calls about method. A Task is something one agent finishes in one session. If it is a task, a direct /goal or a single Kanban card is faster, and define-mission will tell you so. It will never block you. The operator decides; the system advises.
No Lock-In
Any tool that speaks /goal can join a pipeline. There is no proprietary runtime to defend. The Flight Plan compiles down to primitives the ecosystem already understands.