Codeman/agent-teams/experiment-log.md

Agent Teams Experiment Log

Experiment date: 2026-02-12 Test case: ~/claudeman-cases/agent-teams-test/ Team name: research-watchers Teammates: 3 (fs-researcher, perf-researcher, api-researcher) Lead session: 461daa80-94ec-4e5e-a1bb-0518f78311bc Duration: ~3 minutes (06:45:01 → 06:48:07)

Pre-Experiment State

~/.claude/teams/     — did NOT exist
~/.claude/tasks/     — 75 UUID-named directories (from regular Task tool subagents)
Claude processes     — 7 (including watchers)
settings.local.json  — edited to add CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Experiment Prompt

Create an agent team with 3 teammates to research the following topics in parallel:
Teammate 1 fs-researcher researches how Node.js fs.watch works on Linux vs macOS.
Teammate 2 perf-researcher researches inotify performance limits and alternatives.
Teammate 3 api-researcher researches the inotifywait command-line API.
Have each teammate write a brief summary of their findings in a separate file.
Name the team research-watchers.

---

Question 1: What exact filesystem artifacts do agent teams create?

Expected: ~/.claude/teams/research-watchers/config.json and ~/.claude/tasks/research-watchers/

Actual: CONFIRMED + SURPRISE inboxes/ directory

~/.claude/teams/research-watchers/
├── config.json                          # Team config (members, lead, metadata)
└── inboxes/                             # Filesystem-based messaging!
    ├── api-researcher.json              # Per-teammate inbox
    ├── fs-researcher.json
    ├── perf-researcher.json
    └── team-lead.json                   # Lead also has an inbox

~/.claude/tasks/research-watchers/
├── .lock                                # Empty file (presence = lock indicator?)
├── 1.json                               # Task: Research Node.js fs.watch
├── 2.json                               # Task: Research inotify performance
├── 3.json                               # Task: Research inotifywait CLI
├── 4.json                               # Internal: fs-researcher spawn tracking
├── 5.json                               # Internal: perf-researcher spawn tracking
└── 6.json                               # Internal: api-researcher spawn tracking

Subagent transcripts also appear in the standard subagent directory:

~/.claude/projects/-home-arkon-claudeman-cases-agent-teams-test/
└── 461daa80.../
    ├── 461daa80...jsonl                 # Lead session transcript
    └── subagents/
        ├── agent-ae50544.jsonl          # Teammate: fs-researcher
        ├── agent-aa20c65.jsonl          # Teammate: perf-researcher
        ├── agent-a29de32.jsonl          # Teammate: api-researcher
        ├── agent-a04968e.jsonl          # Sub-subagent (teammate's Task tool)
        ├── agent-a0d372e.jsonl          # Sub-subagent
        ├── agent-a2ff939.jsonl          # Sub-subagent
        ├── agent-a89ad82.jsonl          # Sub-subagent
        ├── agent-aa1efc7.jsonl          # Sub-subagent
        └── agent-ab0ef07.jsonl          # Sub-subagent

Cleanup: At 06:48:02, the lead deleted ALL artifacts — inboxes, config, tasks, the team directory itself. Clean removal.

---

Question 2: Is the mailbox/communication filesystem-based or tool-based?

Expected: Tool-based (SendMessage tool), NOT filesystem

Actual: BOTH! Hybrid — tool triggers filesystem writes.

Communication uses the SendMessage tool internally, but the actual message delivery is via filesystem inbox files. Each teammate has ~/.claude/teams/{name}/inboxes/{teammate}.json containing a JSON array of messages.

Inbox message format:

[
  {
    "from": "team-lead",
    "text": "{\"type\":\"task_assignment\",\"taskId\":\"1\",\"subject\":\"Research Node.js fs.watch...\",\"assignedBy\":\"team-lead\",\"timestamp\":\"...\"}",
    "timestamp": "2026-02-12T05:45:18.176Z",
    "read": false
  }
]

Key observations:

• text field is a JSON string (double-encoded) containing a typed message object
• Message types observed: task_assignment, shutdown_request, shutdown_response
• read field tracks whether teammate has processed the message (false → true)
• File locking via .json.lock directories (mkdir-based atomic lock, created then deleted)
• Lead also has an inbox (team-lead.json) for receiving messages FROM teammates

Implication for Claudeman: We CAN intercept messages by watching inbox JSON files! We can also potentially inject messages by writing to inbox files.

---

Question 3: Do teammates appear in the subagents directory?

Expected: Unclear

Actual: YES! Teammates appear as standard subagents.

Teammates create transcript files at:

~/.claude/projects/{hash}/{leadSessionId}/subagents/agent-{agentId}.jsonl

This is the exact same path pattern that regular Task tool subagents use. The existing subagent-watcher.ts successfully discovers them.

Claudeman's /api/subagents endpoint returned them with status "active":

Agent: ae50544  Status: active  Tools: 8   Model: claude-opus-4-6
  Desc: <teammate-message teammate_id= team
Agent: aa20c65  Status: active  Tools: 9   Model: claude-opus-4-6
  Desc: <teammate-message teammate_id= team
Agent: a29de32  Status: active  Tools: 7   Model: claude-opus-4-6
  Desc: <teammate-message teammate_id= team

Distinguishing teammates from regular subagents:

• Description starts with <teammate-message teammate_id= team (a unique marker)
• We can also cross-reference with ~/.claude/teams/{name}/config.json members list

Sub-subagents: Teammates can spawn their own Task tool subagents. 3 teammates spawned 6 additional subagent files (9 total in the subagents directory).

---

Question 4: What does config.json actually look like?

Actual config.json (with all 3 teammates):

{
    "name": "research-watchers",
    "description": "Research team investigating file watching mechanisms...",
    "createdAt": 1770875105373,
    "leadAgentId": "team-lead@research-watchers",
    "leadSessionId": "461daa80-94ec-4e5e-a1bb-0518f78311bc",
    "members": [
        {
            "agentId": "team-lead@research-watchers",
            "name": "team-lead",
            "agentType": "team-lead",
            "model": "claude-opus-4-6",
            "joinedAt": 1770875105373,
            "tmuxPaneId": "",
            "cwd": "/home/arkon/claudeman-cases/agent-teams-test",
            "subscriptions": []
        },
        {
            "agentId": "fs-researcher@research-watchers",
            "name": "fs-researcher",
            "agentType": "general-purpose",
            "model": "claude-opus-4-6",
            "prompt": "You are \"fs-researcher\" on the \"research-watchers\" team...",
            "color": "blue",
            "planModeRequired": false,
            "joinedAt": 1770875126680,
            "tmuxPaneId": "in-process",
            "cwd": "/home/arkon/claudeman-cases/agent-teams-test",
            "subscriptions": [],
            "backendType": "in-process"
        },
        {
            "agentId": "perf-researcher@research-watchers",
            "name": "perf-researcher",
            "agentType": "general-purpose",
            "model": "claude-opus-4-6",
            "prompt": "...",
            "color": "green",
            "planModeRequired": false,
            "joinedAt": 1770875130344,
            "tmuxPaneId": "in-process",
            "cwd": "/home/arkon/claudeman-cases/agent-teams-test",
            "subscriptions": [],
            "backendType": "in-process"
        },
        {
            "agentId": "api-researcher@research-watchers",
            "name": "api-researcher",
            "agentType": "general-purpose",
            "model": "claude-opus-4-6",
            "prompt": "...",
            "color": "yellow",
            "planModeRequired": false,
            "joinedAt": 1770875134997,
            "tmuxPaneId": "in-process",
            "cwd": "/home/arkon/claudeman-cases/agent-teams-test",
            "subscriptions": [],
            "backendType": "in-process"
        }
    ]
}

Key fields per member:

• agentId: {name}@{teamName} format
• agentType: "team-lead" for lead, "general-purpose" for teammates
• model: Model used (inherits from lead)
• prompt: Full spawn prompt (only for teammates)
• color: UI color assignment (blue, green, yellow)
• backendType: "in-process" for in-process mode
• tmuxPaneId: "in-process" or actual pane ID for tmux mode
• subscriptions: Empty array (possibly for message routing)

Config grows incrementally — starts with just lead member (620 bytes), grows as teammates are added (→ 1886 → 3188 → 4551 bytes).

---

Question 5: How do shared tasks differ from regular tasks?

Expected: Team name directory vs UUID, richer task format

Actual: CONFIRMED

Team tasks (~/.claude/tasks/research-watchers/):

{
    "id": "1",
    "subject": "Research Node.js fs.watch on Linux vs macOS",
    "description": "Research how Node.js fs.watch works differently...",
    "activeForm": "Researching Node.js fs.watch Linux vs macOS",
    "status": "in_progress",
    "blocks": [],
    "blockedBy": [],
    "owner": "fs-researcher"
}

Internal teammate tracking tasks (4.json, 5.json, 6.json):

{
    "id": "4",
    "subject": "fs-researcher",
    "description": "You are \"fs-researcher\" on the \"research-watchers\" team...",
    "status": "in_progress",
    "blocks": [],
    "blockedBy": [],
    "metadata": { "_internal": true }
}

Key differences from regular subagent tasks (~/.claude/tasks/{UUID}/):

Feature	Regular tasks	Team tasks
Directory name	UUID	Human-readable team name
File names	.lock, .highwatermark only	Numbered JSON files (1.json, 2.json...)
Content	Lock files only (no task JSON)	Full task JSON with metadata
Owner field	N/A	Teammate name
Locking	.lock file	.lock.lock directory (mkdir atomic)
Internal tasks	None	_internal: true for teammate spawn tracking

---

Question 6: Can we write to task/mailbox files to interact with teammates?

Expected: Possibly for tasks, no for messages

Actual: LIKELY YES for both

Evidence supporting external writes:

1. Inbox files are plain JSON arrays — we could append messages
2. Task files are plain JSON — we could modify status, add new tasks
3. File locking uses .json.lock directories — we'd need to respect the locking protocol
4. Lock protocol: Create directory {file}.lock → write → delete directory. Simple mkdir-based atomic lock.

Not tested in this experiment — would need a follow-up test to verify teammates actually pick up externally-added messages/tasks. But the format is clear and the locking is simple.

---

Question 7: What happens to Claudeman's idle detection with active teammates?

Expected: Lead may appear idle while teammates work

Actual: Lead stays "idle" in Claudeman's view, but terminal shows active status

Observations:

• Claudeman session status showed "idle" throughout the experiment
• The terminal output continued updating (task list checkboxes, teammate progress messages)
• Lead displayed "Befuddling..." spinner while waiting for teammates
• Token count climbed from 27k → 33k during the experiment
• The stop hook DID fire at the end when the team was cleaned up

Implication: Current idle detection may trigger prematurely if:

• It only checks Claudeman's session status (which stays "idle")
• It doesn't account for active teammates

What we need: Check ~/.claude/teams/*/config.json for active members before declaring idle.

---

Question 8: How many Claude processes spawn per teammate?

Expected: 1 claude process per teammate

Actual: ZERO separate processes! Teammates are in-process threads.

# Only 2 claude processes (both Claudeman sessions, none for teammates):
25405 claude --dangerously-skip-permissions --session-id 236f004f...  (our main session)
383633 claude --dangerously-skip-permissions --session-id 461daa80... (test session + 3 teammates)

# Process tree for test session:
claude(383633)─┬─{claude}(383635)
               ├─{claude}(383636)
               ├─... (22 threads total)
               └─{claude}(399252)

In-process mode = threads, not processes. All 3 teammates run as threads within the single claude process (PID 383633). This explains:

• No separate PIDs to track
• No /proc/{pid}/environ for individual teammates
• Lower resource overhead
• Shared env vars automatically

---

Question 9: Do teammates inherit Claudeman env vars (hook events)?

Expected: Yes, if child processes

Actual: YES, trivially — they're in-process threads

Since teammates are threads in the lead's process (PID 383633), they share the exact same environment:

CLAUDEMAN_SCREEN=1
CLAUDEMAN_SESSION_ID=461daa80-94ec-4e5e-a1bb-0518f78311bc
CLAUDEMAN_SCREEN_NAME=claudeman-461daa80
CLAUDEMAN_API_URL=http://localhost:3000

The CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 env var was set via settings.local.json's env key, which Claude Code reads at startup and sets on its process.

Hook events: The lead session's hooks (Notification, Stop) apply to the whole process. Teammate-specific hooks (TeammateIdle, TaskCompleted) are defined in the same settings.local.json and would fire for the lead's session.

---

Question 10: Does subagent-watcher pick up teammates automatically?

Expected: Probably not

Actual: YES! subagent-watcher detects teammates automatically.

Teammates create transcript files in the standard subagent path:

~/.claude/projects/{hash}/{leadSessionId}/subagents/agent-{id}.jsonl

Claudeman's /api/subagents endpoint returned all 3 teammates as active subagents. They're indistinguishable from regular Task tool subagents except:

1. Their description field starts with <teammate-message teammate_id= team
2. They can be cross-referenced with ~/.claude/teams/{name}/config.json
3. They tend to be longer-lived than regular subagents

Sub-subagents: Teammates also spawn their own Task tool subagents (6 additional agents detected), creating a 3-level hierarchy: Lead → Teammates → Sub-subagents.

---

Filesystem Event Timeline

06:45:01  Session transcript created
06:45:05  ~/.claude/teams/ created
06:45:05  ~/.claude/teams/research-watchers/ created
06:45:05  config.json created (lead member only, 620 bytes)
06:45:05  ~/.claude/tasks/research-watchers/ created with .lock
06:45:11  Task 1.json created (via .lock.lock directory lock)
06:45:13  Task 2.json created
06:45:15  Task 3.json created
06:45:18  inboxes/ directory created
06:45:18  fs-researcher.json inbox created (task_assignment message)
06:45:18  perf-researcher.json inbox created
06:45:19  api-researcher.json inbox created
06:45:26  config.json updated (fs-researcher added, 1886 bytes)
06:45:26  Subagent agent-ae50544.jsonl created (fs-researcher)
06:45:26  Task 4.json created (internal: fs-researcher tracking)
06:45:30  config.json updated (perf-researcher added, 3188 bytes)
06:45:30  Subagent agent-aa20c65.jsonl created (perf-researcher)
06:45:30  Task 5.json created (internal: perf-researcher tracking)
06:45:34  config.json updated (api-researcher added, 4551 bytes)
06:45:34  Task 6.json created (internal: api-researcher tracking)
06:45:35  Subagent agent-a29de32.jsonl created (api-researcher)
06:45:35+ Teammates working, additional subagent transcripts appearing
06:47:xx  Tasks completed, shutdown_requests sent to teammate inboxes
06:47:56  team-lead.json inbox created (teammates reporting back)
06:47:57  config.json updated multiple times (member removal?)
06:48:02  CLEANUP: all inbox files deleted
06:48:02  CLEANUP: inboxes/ directory deleted
06:48:02  CLEANUP: config.json deleted
06:48:02  CLEANUP: research-watchers team directory deleted
06:48:02  CLEANUP: all task files deleted (1-6.json + .lock)
06:48:02  CLEANUP: research-watchers task directory deleted

---

Web UI Observations

Terminal output:

• Task list appears with checkboxes: ☐ Research Node.js fs.watch on Linux vs macOS
• Checkboxes fill in as tasks complete: ☑ Research Node.js fs.watch...
• Each task shows assigned teammate: (@fs-researcher)
• Spinner shows active teammate with progress

Status bar:

• Shows team member selector: @main @api-researcher @fs-researcher @perf-researcher
• Hint: shift+↑ to expand and ctrl+t to show teammates
• Standard bypass permissions and token count still visible

Subagent floating windows:

• Teammates DID appear as subagent floating windows in Claudeman's web UI
• They show the standard subagent info (model, tool calls, description)
• Sub-subagents (teammates' own Task tool usage) also appear

In-process mode specifics:

• No new terminal windows or panes
• Everything renders in the single terminal session
• Shift+Up/Down would switch between teammate views (not tested interactively)

---

Conclusions & Key Surprises

Surprises vs expectations

1. Inboxes ARE filesystem-based — contrary to docs saying "SendMessage tool". It's a hybrid: the tool writes to filesystem inboxes.
2. Teammates are threads, not processes — no new OS processes, just threads within the lead's claude process.
3. Teammates appear as standard subagents — existing subagent-watcher infrastructure works out of the box!
4. Config grows incrementally — members are added one-by-one, not all at once.
5. Internal tracking tasks — tasks 4-6 with _internal: true track teammate spawn state.
6. Auto-cleanup — lead automatically cleaned up ALL artifacts after shutdown.
7. Sub-subagents — teammates can spawn their own Task tool subagents (3-level hierarchy).
8. teammateMode is NOT a valid settings key — display mode defaults to in-process.

Design implications for Claudeman

1. TeamWatcher can be simple — just poll ~/.claude/teams/ for directories + parse config.json
2. Subagent-watcher already works — no new infrastructure needed for teammate transcript tailing
3. Idle detection needs team awareness — check config.json members before declaring idle
4. Message interception is possible — watch inbox JSON files for real-time message tracking
5. Task visualization is straightforward — parse numbered JSON files in task directory
6. No process tracking needed — teammates are threads, not separate processes
7. Distinguish teammates from subagents — use description prefix <teammate-message or cross-reference config.json

What to build first

1. Team-aware idle detection — highest priority, prevents premature respawn
2. TeamWatcher — poll ~/.claude/teams/ for team creation/removal
3. Team tasks API — parse task JSON files for UI display
4. Teammate badge in subagent windows — mark teammate subagents differently from regular ones
5. Message timeline — parse inbox files for inter-teammate communication display

What we DON'T need to build

• Process discovery for teammates (they're threads)
• Custom transcript tailing (subagent-watcher handles it)
• Separate teammate window infrastructure (subagent windows work)