Skip to content

Team Coordination

li team gives agents in a flow a persistent inbox for mid-run signals. The team file survives at ~/.lionagi/teams/<id>.json after the flow exits.

Setup

pip install lionagi          # or: uv add lionagi
# claude — Option A (subscription): npm install -g @anthropic-ai/claude-code && claude login
#          Option B (API key):      export ANTHROPIC_API_KEY="sk-ant-..."

Create a team

li team create "research-review" -m researcher,reviewer,orchestrator

# output:
Created team 'research-review' (e7a3d91bf542)
  Members: researcher, reviewer, orchestrator
  File: ~/.lionagi/teams/e7a3d91bf542.json

Run a flow with team mode

li o flow claude/sonnet \
  "Research Python async patterns, then review the draft for completeness" \
  --team-mode research-review --save ~/team-out

# output:
Team 'research-review' created (c4f8b2a01e73)
  ▶ researcher started
  ✓ researcher done (14.2s)
  ▶ reviewer started
  ✓ reviewer done (9.1s)
Saved to /home/user/team-out
[...agent results...]

--team-mode research-review creates a fresh team per run with that name and injects TEAM_COORD_SECTION into each worker's system prompt.

Mid-flow signals

Workers call li team send while their op is still running:

li team send \
  "Draft at ~/team-out/agents/researcher/research.md — ready for review" \
  -t research-review --to reviewer --from researcher --from-op o1

# output:
Sent to reviewer in 'research-review'

The reviewer reads its inbox before starting its own op:

li team receive -t research-review --as reviewer

# output:
[2026-04-20T14:03:08] researcher op=o1 → reviewer
  Draft at ~/team-out/agents/researcher/research.md — ready for review

(1 message)

--as reviewer marks the message read for that member only. Concurrent sends serialize under fcntl.flock — parallel workers don't clobber each other.

Inspect the inbox

After the flow, check the full thread:

li team show research-review

# output:
Team: research-review (c4f8b2a01e73)
Created: 2026-04-20T14:02:55.123456+00:00
Members: orchestrator, researcher, reviewer

────────────────────────────────────────────────────────────
  [2026-04-20T14:03:08] researcher op=o1 → reviewer  (read by: reviewer)
    Draft at ~/team-out/agents/researcher/research.md — ready for review

  [2026-04-20T14:04:31] reviewer op=o2 → *
    Review complete. 2 gaps noted in ~/team-out/agents/reviewer/review.md

Run li team list to see all teams.

How it works

--team-mode research-review creates a team at flow start and injects TEAM_COORD_SECTION into each worker's system prompt — the agent knows its team id, role, and teammates' names. Workers call li team send --from-op <id> to tag which op the signal belongs to; one agent can run several ops on the same branch and each message stays traceable. li team receive --as <name> returns only unread messages for that member and timestamps them as read. At flow end, each worker's result is also posted as a team message, so li team show shows mid-flow signals and final outputs in one thread.

Next