Mission Gradient Explained
A normal goal tells an agent where to end. A mission gradient tells it how to keep moving uphill when the path changes.
Mission gradient is a way to guide long-running agent work without reducing the work to a brittle checklist.
A short task can be handled by instructions like:
Add background agent support.But a long run eventually hits ambiguity. The original checklist becomes stale. The agent discovers unexpected failures, tests contradict assumptions, local fixes create global risks, or the "obvious" implementation starts faking the hard part.
A mission gradient gives the agent a control surface instead of just a destination.
It says:
Move this system in this direction.
Preserve these invariants.
Treat this evidence as real proof.
Prefer moves that improve the artifact's long-term shape.
When the plan becomes wrong, update your belief state instead of pretending the checklist is still enough.At sixth-grade level: a normal goal says, "Get to that mountain." A mission gradient gives you a compass, a map, and rules for what counts as getting closer: keep walking where the land rises, don't cross cliffs, don't leave your team behind, and if the trail disappears, stop and figure out where you are before guessing.
For CLI-native agent work, mission gradient is like having an experienced sysadmin at your shoulder:
Before you run that, check what machine you're on, what user you are, what files it touches, what backup exists, what success looks like, and what you'll do if it half-fails.
It is not slower. It prevents fake speed.
The pieces
A mission gradient should name:
- **Real artifact** — what durable thing should exist at the end?
- **Invariants** — what must not break?
- **Value criterion** — what counts as uphill?
- **Belief state** — what does the agent believe, and why?
- **Homotopy parameters** — what can deform from low-resolution to high-resolution without faking topology?
- **Dense feedback** — what tests, traces, screenshots, logs, or user-visible artifacts keep the run in contact with reality?
- **Anti-Goodhart constraints** — how might the agent satisfy the metric while violating the mission?
- **Rollback policy** — how can the run retreat without losing unrelated good work?
- **Stopping rule** — when should it stop, narrow, or ask rather than continue into slop?
Why it matters
Most agent prompts are written as task lists. The agent tries to finish the list. That works for twenty minutes. It fails over hours because the environment changes faster than the prompt.
Mission gradient tells the agent what "better" means during the whole run. It makes long-running work more like steering and less like executing a shell script copied from the internet.
One sentence
- Mission gradient is a way to prompt agents so they optimize toward a durable direction instead of merely completing a brittle checklist.**