Codename pattern

Alfred-OS expects one agent script per narrow specialist, named after a coherent fictional cast, coordinating via labels + GitHub state, not in-process calls.

What “narrow specialist” means

One codename agent, one job, 150-300 lines:

Codename (Batman cast)	Single job
Lucius (Lucius Fox)	Pick the oldest `agent:implement` issue, claim it, ask Claude to implement, push a branch, open a PR.
Drake (Tim Drake)	Read specs + roadmap + code-reality grep, file the next well-scoped `agent:implement` issue.
Bane	Pick the lowest-coverage actively-changed file, write tests, open a PR.
Ra’s al Ghul	Multi-axis review on every fresh PR.
Nightwing	Apply P0/P1 reviewer comments on `agent:authored` PRs.
Robin	Triage new bug-report issues; classify severity, ask for repro info.
Huntress	Post-deploy E2E smoke against staging.
Gordon	Daily ECS drift + Sentry top-N read.
Alfred	Cross-repo coordinator for features that span multiple repos.
Bat-Signal	Slack notifier for the other agents.

What the pattern is not:

Not “one agent does everything”. A single Lucius doing feature dev + tests + review + triage + smoke would be unmaintainable, prompt-engineering hell.
Not “the smallest possible unit of work per agent”. A separate codename for “create branch”, “commit”, “push” is surgery for the sake of it.

The right granularity is one human role. If you’d hire a junior to do this job and review their work, it’s a codename.

Why a fictional cast

Two reasons.

Operational legibility

Codenames show up in:

PR titles (feat(events): add nullable cost column [Lucius])
Commit trailers (Agent-Codename: lucius)
Slack messages (✅ Lucius shipped: <url>)
Issue labels (lucius-attempt-1, lucius-pr-open)
Worktree paths (~/.hermes/worktrees/eng-lucius-backend-303-...)
Logs (/tmp/my.fleet.lucius.stdout)

If your cast is “agent-1 / agent-2 / agent-3” or “feature-dev / test-coverage / review”, scanning the firehose becomes laborious. A coherent fictional cast (Batman, Greek mythology, The Wire) makes “Lucius failed on #303” instantly readable.

Design forcing function

“What does Bane do?” is a sharper question than “what does the test agent do?”. Naming the role after a character (with a personality, a domain, relationships to other characters) forces you to decide:

What’s Bane’s scope? Brute-force test coverage on changed files. Not unit-test design philosophy.
What does Bane never do? Never modifies non-test files. Never opens an architecture issue.
How does Bane interact with the others? Bane’s PRs go through Ra’s al Ghul like any other PR. Bane consumes from the same agent:implement queue Lucius does, but only files labelled test-coverage.

Without the codename, “the test agent” tends to creep: “well, while it’s there, it could also lint… and run a security scan…”. With the codename, the answer is “no, that’s not Bane. Bane writes tests.”

Pick your own cast

The shipped examples use Batman side-characters because the original operator liked Batman. Pick anything coherent:

Greek pantheon: Athena (planner), Hephaestus (feature dev), Hermes (notifier), Asclepius (deploy health).
The Wire: Bunk (review), McNulty (triage), Omar (security audit), Lester (bug investigation).
Tolkien: Aragorn, Legolas, Gimli, Gandalf. Watch lore consistency (Gandalf shouldn’t review Frodo’s PR).
Your favourite anime, novel, podcast, board game. All work.

Constraints:

ASCII-safe names (used in filenames, label slugs, gh CLI args). rasalghul not Ra's al Ghul.
~10 characters max. Long codenames pollute Slack scrolling.
Pronounceable. The operator is going to say “Lucius shipped #303” out loud at some point.
Consistent across the fleet. Don’t mix Batman + Star Wars; pick one universe.

The wiring

Each codename has:

A bin script: bin/<role>.py. Imports from agent_runner. ~150-300 lines.
A launchd entry: one line in launchd/agents.conf (label, script, schedule, java flag).
(Optional) A prompt file: prompts/<role>.md in this repo or $HERMES_HOME/prompts/<codename>.md in your fleet. Long-form context the runner inlines into claude -p.
(Optional) An IAM identity: if it touches AWS. See AWS setup.
A row in your fleet’s CLAUDE.md documenting role + trigger + scope.

Tutorial for an end-to-end build of one codename. Use docs/AGENTS.md for the default cast map and role boundaries.

Anti-patterns

Generic codenames: “agent-1”, “feature-bot”, “the planner”. The cast disappears as a forcing function; prompts bloat.
Code-named-after-tools: “lucius-grpc”, “bane-pytest”. Couples the codename to the implementation; can’t refactor the tool without renaming the role.
Cross-cast mixing: Lucius (Batman) + Hermes (Greek) + Bunk (The Wire). Chaotic in Slack.
One codename per repo instead of per role: “backend-bot”, “frontend-bot”. Loses the role-as-narrow-specialist forcing function.
Codename as adjective: “smart-lucius”, “fast-lucius”. The codename is the specialist; modifiers don’t add anything.