← All reports · published 2026-04-19

Agent Coordination Map

Every agent in Mark's stack — what it can do, where it runs, how it talks to the others, and what every agent should do in the first 30 seconds of a session.

Why this map exists. Prime Directive #1 is integration + communication. Five agents all touching the same codebase, memory, and stack will step on each other unless each one knows (a) what the others are for, (b) which channel to use, and (c) what state to check at session start. This is that one-pager.

Read this at session start. Before anything else.

Refresh: hand-curated. Update when an agent is added, a channel changes, or the session-start routine changes.

1. Agent capability matrix

Agent	Runs on	Primary role	Has memory MCP?	Can execute code on Apex?	Can execute code on G16?	Best for
G16 Claude Code (that's usually me)	G16 (Windows/WSL2)	Primary operator / architect / orchestrator	✅ HTTPS	via SSH	✅ direct	Planning, cross-cutting changes, PR creation, git ops, MCP registry work
Apex Claude Code	Apex (Windows, persistent session)	On-box implementer where TP3 actually runs	✅	✅ direct	via scp/ssh	Postgres queries, MinIO ops, scheduled-task installs, anything touching the live stack
Cursor	G16 (Remote-SSH to Apex available)	IDE-based coder for iterative / multi-file work	❓ (audit pending)	via Remote-SSH	✅	New features, refactors, multi-file edits, anything that benefits from an IDE
Jules (Google)	Cloud, watches GitHub	PR reviewer + test suggester	❌ (no MCP, sees PRs only)	❌	❌	Reviewing PRs, security audits, code-health passes, test coverage suggestions
Antigravity	G16 (Windows desktop app)	Research agent + Google ecosystem	✅ (registered, UI broken for HTTP MCP)	❌	❌ (file-based bridge only)	Gemini-heavy research, Google Drive / Workspace automation, business-plan content
Gemini CLI	Apex	Sidecar for headless Gemini calls	✅ (`~/.gemini/settings.json`)	✅ via `--approval-mode yolo`	❌	One-shot Gemini prompts inside scripts, embedding calls, YAML/JSON generation
Claude Desktop	G16	Interactive chat + writing assistant	via omi-mcp-public OAuth	❌	❌	Customer-facing prose, proposal editing, final-pass polish
claude.ai web / mobile	cloud / phone	Ambient Claude for Mark on the go	✅ via OAuth gateway	❌	❌	Asking TP3 questions from phone, reading digests, brain-dump capture

Hardware context: Apex = always-on Windows 11 server. G16 = primary dev laptop. Samsung Chromebook + Pixel 8 Pro = mobile. Tailscale mesh connects everything.

2. Channel matrix — "when I need to talk to X, I use Y"

I am	I need to…	Channel	Latency	Notes
Any agent	Read / write shared memory	tp3-omi MCP (`twin_memory_*` tools)	<1 s	Canonical. Everyone has this.
Any agent	Drop work for another machine to execute	`BRIEFS/pending/<slug>.md`	≤5 min (runner polls)	G16 runner exists; Apex runner pending. Use `target: any` for doc-only briefs.
G16 Claude	Tell Apex Claude something non-urgent	Git commit to `main` + a line in `MEMORY.md`	next `git pull`	Passive — Apex picks up on next pull.
G16 Claude	Tell Apex Claude something urgent	Drop a BRIEF OR write via `twin_memory_broadcast`	minutes	Broadcast memories are visible to all agents' next query.
Any agent	Ship code for review	Open a PR against `main`	minutes	Jules auto-reviews. Tag `@MrB-Ed` + any agent you want to pull in.
Any agent	Notify Mark	Slack DM to `U0AK03UP9SQ` via `slack_post_message`	seconds	Keep to ≤10 bullets, no prose paragraphs.
Any scheduled job	Emit a status / heartbeat	ntfy via `tp3_ntfy_notify.py`	seconds	Push to Mark's phone. Use for automated pings, not for inter-agent.
Antigravity	Tell other agents something	Append to `_AGENT_COMMS_LOG.md` in the shared Drive folder	next sync	File-based bridge (Antigravity's MCP UI is broken for HTTP). Newest entry at top, signed.
Any agent	Propose a cross-agent change	Open a GitHub issue → `@Jules` for review	async	Issues are the durable record; chat isn't.

Channel anti-patterns (do not do)

Do not use Slack to coordinate between agents. Slack is for Mark. Inter-agent = memory MCP + BRIEFS + GitHub.
Do not SSH into Apex to restart TP3 services — SSH-launched wrappers die on disconnect. Use the scheduled task (Start-ScheduledTask -TaskName "TP3 Launch At Logon"). See feedback_apex_restart_via_scheduled_task.md.
Do not paste commands into Mark's terminal. Ever. Write a one-click PowerShell script in tp3_scripts/ and tell him BOSS, run <script>.ps1.

3. Session-start bootstrap (every agent does this first)

When a new session starts, before doing anything else, run the following:

Fast path (when `status.thebarnetts.info/tp3.json` lands — PR pending)

curl -s https://status.thebarnetts.info/tp3.json

One call → row counts, ingest worker health, queue depth, pending BRIEFS, git HEAD. Proceed.

Slow path (today, until status JSON ships)

Run these in parallel if you can:

1. git fetch origin main && git log origin/main..HEAD --oneline — am I behind?

2. Read MEMORY.md (should be in your context automatically if you're Claude).

3. ls BRIEFS/pending/ — is there work for me?

4. curl http://127.0.0.1:8944/health (if you're on Apex) or note: "ingest worker health unverified from this agent."

5. Announce yourself: twin_memory_write with tag session_start, body <agent_name> online at <timestamp>.

When the status JSON ships

Update step 1 of the fast path here to also show "generated_at" so agents can detect stale status.

4. Who does what — decision tree

"I have a task to hand off. Which agent?"

Does it need to run code on Apex (Postgres, MinIO, ingest worker)?
├── YES → Apex Claude (BRIEF) or Gemini CLI (for one-shot Gemini calls)
└── NO
    Does it need an IDE / multi-file refactor / iteration?
    ├── YES → Cursor (mission brief in repo root)
    └── NO
        Is it code review / security audit / test coverage?
        ├── YES → Jules (GitHub issue)
        └── NO
            Is it research / Gemini deep-dive / Google ecosystem?
            ├── YES → Antigravity (via _AGENT_COMMS_LOG.md or a Google Docs outline)
            └── NO
                Is it customer-facing prose?
                ├── YES → Claude Desktop (Mark pastes the draft in)
                └── NO → I (G16 Claude) probably just do it directly

5. Known coordination debt

Apex BRIEFS runner not installed — currently only G16 runs the BRIEFS scheduled task. Briefs with target: apex need manual trigger by Apex Claude. Fix planned: port _install_g16.ps1 to _install_apex.ps1.
Cursor MCP config unaudited — we don't know which MCP servers Cursor sees. Tracked via a separate Cursor mission brief.
Antigravity HTTP MCP UI broken — Antigravity can only reach the memory MCP via stdio proxy, which isn't wired yet. Until fixed, Antigravity uses _AGENT_COMMS_LOG.md only.
No auth rotation calendar — X-TP3-MCP-TOKEN, TP3_OMI_API_KEY, Cloudflare API token all have no scheduled rotation. Tracked in credential_map.md.
No dead-man's switch — if every agent silently stops updating their session_start memory, nobody notices until Mark asks. Needs a daily heartbeat check (candidate for the status JSON).

6. Related maps

mcp_server_map.md — which MCP tools live where
data_flow_map.md — how data moves through the stack
scheduled_task_map.md — every scheduled task + health
process_map.md — live process inventory on Apex
credential_map.md — every credential + rotation state
known_deferred_map.md — intentionally half-done things (read before "fixing" anything weird)