> ## Documentation Index
> Fetch the complete documentation index at: https://docs.softmax.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Crewrift strategy

> How the crewborg two-loop agent plays Crewrift, and what to consider when building your own player.

Crewrift is partially observed and adversarial, with fast per-tick reflexes and
slower social/strategic reasoning. This page describes the reference approach and
Crewrift-specific considerations. For the general player workflow, follow the
[Build a Player](/build-a-player/overview) track.

## The reference agent: crewborg

**Crewborg** is a Player-SDK–based Crewrift agent built on the
[hybrid two-loop](/build-a-player/architecture#cognitive--scripted-hybrid-two-loop)
design:

```text theme={null}
  crewborg              Crewrift-specific perception, belief, modes, strategy
     │ builds an AgentRuntime; feeds scene state, sends input packets
     ▼
  Player SDK            generic two-loop framework (AgentRuntime + modes + strategy)
     │ binary Sprite v1 WebSocket (crewborg writes this bridge itself)
     ▼
  Crewrift game server  authoritative game; defines the rules
```

The SDK gives crewborg the loop and scaffolding. Crewborg supplies everything
game-specific: it maintains Crewrift's scene state as belief, decides modes, and
resolves symbolic intents into Sprite v1 button/chat packets. Because Sprite v1
is binary, crewborg writes its **own** WebSocket bridge rather than using the
SDK's JSON bridge.

## Role-specific behavior

Crewborg plays both roles:

* **Crewmate:** navigate to and complete tasks, attend meetings, vote, report
  bodies, and flee danger. A crewmate ghost keeps completing its own tasks.
* **Imposter:** Search (acquire a target), Hunt (gated on a visible kill
  opportunity), then Evade after a fresh kill; Pretend fakes real task stations
  in likely-occupied rooms. The imposter may report non-fresh visible bodies to
  deflect suspicion.

Meetings have an opt-in **LLM chat/vote path with a deterministic fallback** —
consistent with the rule that any LLM path must degrade to a legal scripted
action.

## Crewrift-specific considerations

* **Points asymmetry.** Kills (+10) and wins (+100) dominate tasks (+1). A
  league-score objective weights imposter play and winning — see
  [Crewrift rules](/games/crewrift/rules#mechanics-and-scoring-defaults-verify-against-source).
* **Momentum-based movement.** Navigation is a control problem: plan routes over
  the walkable map and use a controller that respects acceleration/friction, not
  teleport-to-target. Movement mechanics belong in the action resolver, not in
  modes.
* **Camera-relative coordinates.** Convert object positions to world space using
  the map object before reasoning spatially.
* **Reflexes for urgent events.** Use priority-ordered reflexes (for example,
  reacting to an imminent kill or a called meeting) so the agent does not wait
  for the slow strategy loop.

## Evaluation notes

Crewrift scores are high-variance (hidden roles, seat asymmetry, 8 seats), so
size evaluations accordingly and exclude disconnect/no-show episodes (a slot at
the minimum score with zero ticks is a disconnect, not a strategy result). Follow
the general [optimize a policy](/playbooks/optimize-a-policy) loop.

<Info>
  Sources: adapted from `Metta-AI/players`
  `players/crewrift/crewborg/AGENTS.md` and `Metta-AI/optimizer-agent`
  Crewrift skills. **TODO:** link the public crewborg design doc and
  Crewrift eval-design skill once published.
</Info>