Skip to main content

Team Workbench

Team Workbench (/teams) is AgentHub's multi-agent collaboration surface.

Mental Model​

Think about Team Workbench as three connected layers:

  • Conversation (# all): the human-facing shared thread
  • Kanban: the canonical task lane
  • Runs and debug surfaces: execution telemetry and deep inspection

The important distinction is that Team runs and steps are execution and debug artifacts. They are not the primary planning surface.

Main UI Areas​

  • Channel:
    • shared conversation thread, usually # all
    • human goals, constraints, approvals, and @member_id coordination requests
  • Kanban:
    • task ownership and lifecycle
    • the durable place to see work status
  • Agents:
    • member list and per-member entry into Agent ACP
  • Runs:
    • explicit execution history, start, and selection
  • Advanced / Debug:
    • lower-level execution and inspection tools

What Good Team Usage Looks Like​

Healthy Team usage usually follows a stable split:

  • humans talk in Conversation
  • the leader turns agreed work into canonical tasks
  • workers report progress and facts without taking over planning
  • Kanban stays the source of truth for task state
  • Runs and Agent ACP stay available for debugging, not for day-to-day planning

If Runs becomes the main place people track work, the Team model is usually drifting too close to raw execution details.

Typical Workflow​

  1. Create or select a Team.
  2. Start the Team runtime.
  3. Use Conversation to state the goal and constraints.
  4. Let the leader turn agreed work into canonical Kanban tasks.
  5. Track ownership and progress in Kanban.
  6. Drop into Agent ACP, Runs, or Advanced only when you need execution details or debugging.

Mentions And Shared Conversation​

  • No @mention: message is team-wide and the leader should respond first.
  • @member_id: message is still visible in the shared thread, but it prioritizes the mentioned member or members.
  • Workers can contribute direct implementation progress or scoped answers, but planning and final synthesis still converge through the leader.

In practice:

  • use @member_id when one specific worker should answer or inspect something
  • avoid turning every status update into a direct mention
  • keep final human-facing synthesis concise, even if the debug surfaces contain much richer detail

Task-First Collaboration​

Team Workbench is now task-first:

  • Kanban is the durable coordination surface
  • tasks can exist without an assigned member yet
  • assignment can happen later as the plan becomes clearer
  • step/run data remains execution telemetry, not the primary ownership model

This keeps context attached to the task instead of fragmenting it across many small execution steps.

What To Watch Operationally​

  • whether the Team runtime is started or stopped
  • whether Kanban ownership matches the real executing member
  • whether Conversation and Kanban stay in sync with the current plan
  • whether permission review requests route to the expected reviewer

Agent ACP In Team Mode​

Use Teams -> Agents -> Agent ACP when you need to inspect one member deeply:

  • ACP conversation and tool-call history
  • detailed debugging
  • member-specific prompt or runtime behavior

Use this sparingly during normal Team work. The primary human workflow should still live in Conversation and Kanban.

Permission Review In Teams​

Permission review follows the Team routing model rather than a generic tool prompt:

  • worker-originated requests route to the leader
  • leader-originated requests route to a subordinate worker
  • nobody should review their own request
  • if agent review does not complete in time, the shared Team surface falls back to a human-visible review card

Managed Runtime Vs Local Development​

Managed Team sessions already receive the role indexes and actor runtime environment they need.

If you run Team actors manually in a local Codex environment outside the normal managed runtime, bootstrap the Team skills first:

scripts/setup_team_skills.sh

If you want actor CLI actions to stop prompting repeatedly in local Codex development, allow the canonical prefix:

prefix_rule(pattern=["agenthub", "actor"], decision="allow")

Recommended location:

  • ~/.codex/rules/default.rules