Cursor rule

1---
2description: "MiniMax M3 + Cursor 3.7 agent teams: role boundaries, multi-environment handoffs, /best-of-n as a team pattern, escalation, and serial vs parallel team structure."
3alwaysApply: false
4---
5 
6# Agent Teams (M3 + Cursor 3.7)
7 
8Use this rule when the work is best handled by multiple agents or explicitly separated roles.
9 
10## Team Shape
11 
12- Define roles before delegating: planner, explorer, implementer, verifier, or another bounded role.
13- Give each role a distinct objective, owned surface, and stopping point.
14- Do not create overlapping ownership without a clear reason.
15 
16## Multiple concurrent Cursor sessions
17 
18Cursor 3.7 can run **several agent sessions in parallel** (for example from the Agents Window sidebar, or kicked off from mobile, web, Slack, GitHub, Linear). The same rules apply as for parallel roles or subagent branches: use parallel sessions only for **independent** workstreams, keep **non-overlapping ownership** of files and decisions, and **serialize** anything that would race the same artifact. More sessions visible at once does not make conflicting edits safer.
19 
20## Multi-Environment Teams
21 
22When a single team spans environments (some roles local, some in `/worktree`, some in cloud, some over SSH):
23 
24- The handoff contract must name the **environment** per role — local, worktree, cloud, or SSH — and the **model** per role.
25- Choose environments to remove file-system races: if two roles may touch the same artifact, pin both to the same isolated worktree or serialize the touch.
26- Prefer `/worktree` for any role whose output will be reviewed and possibly discarded. Prefer `cloud` for any role whose value is "keep working while I am away."
27- When one role's output feeds another role in a different environment, the handoff payload should include the file paths, branch, and any required state — not just a prose summary.
28 
29## Handoff Contract
30 
31Every handoff should state:
32 
33- current goal
34- **which repo or workspace root** when more than one is in play
35- **environment** (local / worktree / cloud / SSH) and **model** (e.g. `MiniMax-M3`, `composer-2.5`, `auto`)
36- files, dirs, or questions owned
37- findings or artifacts produced
38- open risks or assumptions
39- exact next step for the receiving role
40 
41## Serial vs Parallel
42 
43- Use parallel teams only for independent branches.
44- Use serial teams when one role depends on another role's findings.
45- If two agents may touch the same file or decision, centralize that step instead of racing them.
46 
47## /best-of-n as a Team Pattern
48 
49For decisions where a second opinion is cheap insurance (architecture choice, design parity, risky refactor, choosing between two valid approaches), treat `/best-of-n` as a team pattern:
50 
51```text
521. Define the question and the deliverable shape
532. Run the same prompt across 2-4 models in parallel worktrees
543. Read each result centrally
554. Pick the strongest, merge complementary ideas, or escalate to the user with the comparison
56```
57 
58The "team" is the synthesis step, not the model run; centralize it.
59 
60## Challenge And Review
61 
62- At least one role should challenge assumptions, edge cases, or weak evidence on non-trivial work.
63- Reviewer or verifier roles should critique the work, not merely restate it.
64- Keep disagreement evidence-based and scoped to the task.
65- A verifier reviewing visual work should re-read the actual post-change frame, not rely on a prose description; that is `multimodal-grounded` review.
66 
67## Human Escalation
68 
69Escalate to the user when:
70- a branch changes architecture or product direction
71- safety, destructive actions, or irreversible changes appear
72- two valid team recommendations lead to meaningfully different outcomes
73- `/best-of-n` returns a near-tie that needs human taste
74 
75## Closeout
76 
77- Synthesize in the main thread before acting on team output.
78- Final user-facing claims must follow the always-on status and verification contract.
79- Include the environment, the model, and the relevant evidence (including `multimodal-grounded` for visual claims) in the closeout.
80