Agentic Government — a Claude Code skill

Run Claude Code like a small, well-run government. One manager model talks to you and plans the work. It delegates each piece to a cheaper or stronger worker model according to how hard the task is. A top-tier checker verifies the result before anything counts as done. A ledger records the cost, time, checker verdicts and rework of every job, so you can see whether the whole arrangement actually saves you tokens and time.

By Alex Glonti · Nomad Devs (nomaddevs.club)

Status: v0.1 — extracted from a few days of real use on a two-person studio's release work. The structure is stable; the numbers below are early.

The roles

Each worker is a Claude Code agent file (agents/*.md) with a pinned model and a restricted tool list. The manager is your own session.

Role	Agent file	Model	Does	Never does
Zeus, the Chancellor (manager)	your session	Opus or top tier	talks to you, plans, delegates, decides, appends the ledger	long builds, watcher loops
Pip, the Clerk (runner)	runner	Haiku	read-only chores: find files, grep, read logs, watch a folder	edit, build, send, decide
Iris, the Envoy (researcher)	researcher	Sonnet	web research, policy checks, comparisons, drafts into a file	publish, send, change accounts
Gela, the Engineer (builder)	builder	Opus	real changes: code, scripts, assets, release prep	irreversible or outward actions
Vera, the Auditor (checker)	checker	top tier (Fable)	read-only audit: PASS or FIX-NEEDED with evidence	fix the work herself
Clio, the Chronicler	chronicler	Haiku	the ledger and its reports	project work
Hestia, the Housekeeper	housekeeper	Sonnet	finds clutter, moves it to a dated _recycle folder with a manifest; lists running tools	delete anything
Dionysus, the Fool	muse	Sonnet	three wild ideas per brief (odd, ambitious, crazy)	act on them
Metis, the Treasurer	treasurer	Sonnet	efficiency audits from the ledger and over-budget transcripts	stop anyone mid-job

If the top tier is scarce (plan limit, budget), change model: fable to model: opus in agents/checker.md. The checker stays a separate, read-only reviewer, just on a lower tier; note the tier on each verdict. We ran this way for part of one week.

The rules that keep it from collapsing

Most multi-agent setups stall the same way: nobody in charge, nobody checking, cheap bots doing things nobody can undo.

1. Check by stakes. Trivial read-only chores are glanced at by the manager. Anything that goes public, into a build, to a store, or is reported as "done" goes through the checker first. A FIX-NEEDED goes back to the same worker, then is re-checked once.
2. Irreversible steps are human-only. Publishing, submitting, buying, sending, deleting, pushing, store and account changes are done only by the manager, and only with your explicit OK. A check after the fact cannot undo them.
3. Final screens are human-only. On the last confirmation page of a store, bank or publishing service, the agent touches nothing, not even to open a panel, so there is never doubt about who confirmed.
4. Tools stay dormant. Servers, models, editors and watchers are not started unless the job needs them; whoever starts one stops it. At the end of the day the housekeeper lists what is still running.
5. Two failures mean switch. Two failed rounds on the same defect: change the tool or the tier, do not try a third time the same way. Every brief carries a budget ("at 70%, stop and report").

Real-world example (generic)

A two-person game studio used this setup to ship a mobile game update over a weekend, on Android and iOS, with the developers making the decisions and the final clicks. Counting by hand from the first three days of its ledger:

• About 57 logged jobs in three days: builds, store-listing and website drafts, art and audio passes, research, clean-ups, media cuts.
• 12 jobs went through the checker with a PASS, and 5 jobs drew at least one FIX-NEEDED (8 FIX-NEEDED verdicts in total). Catches included confidential wording in a draft public text, a build number the store would have rejected, a draft that contradicted the real product, and image faults.
• Rework rounds the human had to request: low, mostly 0 or 1 per job. The heaviest cases were creative (an image redo needed two). The ledger counts a checker's FIX-NEEDED as rework too, so the numbers cannot hide behind the human's patience.
• A baseline script (tools/measure_sessions.py) compares sessions before and after adoption: tokens per user turn, rework turns, tool errors, delegated runs.

These are one team's numbers over a short window. Treat them as an example of what the ledger shows, not as a benchmark. The honest result so far is "fewer defects reach the human, and every claim has evidence", not a proven saving.

Quick start

1. Copy into your Claude Code user folder (~/.claude/ on macOS and Linux, C:\Users\<you>\.claude\ on Windows):

   agents/*.md                    -> ~/.claude/agents/
   skills/agentic-government      -> ~/.claude/skills/
   

2. Copy ledger-template.md somewhere you keep notes. The manager appends one line per job; Clio turns it into reports.

3. Start a new Claude Code session and say "use the government" (or run /agentic-government) on any multi-part job.

4. Optional, read-only baseline of your own sessions: python tools/measure_sessions.py --help.

If your plan lacks a model, change the model: line in that agent file.

Token budget / context watchdog

Every reply re-sends the whole conversation, so cost grows with length: in a long agent day about 99.9% of tokens were cache reads of re-sent context, and a turn at 900K context costs roughly 4x one at 250K. The rules:

• Builders hand back at ~300K context or 150 turns with a state note (done, left, next step); a fresh agent continues.
• Tuning loops get at most 3 tries, then report.
• Use the strongest model only where the job needs it.
• When the manager's own context gets long it writes a memory handoff and restarts.
• hooks/context_watchdog.py warns you (250K, then 450K, then every +150K) and logs it.

Install: copy hooks/ to ~/.claude/hooks/ and add to ~/.claude/settings.json (use python3 on macOS/Linux; on Windows use the full path to python.exe if python is not on PATH):

{
  "hooks": {
    "Stop":         [{"hooks": [{"type": "command", "command": "python ~/.claude/hooks/context_watchdog.py", "timeout": 5}]}],
    "SubagentStop": [{"hooks": [{"type": "command", "command": "python ~/.claude/hooks/context_watchdog.py", "timeout": 5}]}]
  }
}

Thresholds and the log path live in context_watchdog.json (default log ~/.claude/hooks/token_watch.log; the AG_TOKEN_LOG environment variable overrides it).

When not to use it

A quick question or a one-file edit: just ask normally. Delegation has overhead and pays off on jobs with several independent pieces or long chores.

What is inside

• skills/agentic-government/SKILL.md — the manager's playbook: tiers, how to run a job, when not to delegate, the efficiency rules
• agents/ — the eight worker files (runner, researcher, builder, checker, chronicler, housekeeper, muse, treasurer)
• ledger-template.md — the ledger columns
• hooks/context_watchdog.py — Stop/SubagentStop hook that warns when context gets long
• tools/measure_sessions.py — read-only session metrics (numbers only, no message text)

Licence

MIT. See LICENSE.

By Alex Glonti · Nomad Devs (nomaddevs.club)