Agent roster

July 7, 2026 · View on GitHub

The named specialists the driver can delegate to. Each is a plain markdown file in claude/agents/name, a description the driver reads to auto-pick the right one, a deliberately-chosen model tier, and a scoped tool set. The description is the routing signal: it says when to reach for the agent, so the driver delegates without being told which one.

Model tier follows the cost discipline in claude/CLAUDE.md: cheap models for mechanical/parallel work, the strong model only where a wrong answer is expensive. Reserve Opus for judgment; push fan-out down to Sonnet and Haiku.

  • Haiku — mechanical, high-volume: test runs, log triage, formatting sweeps.
  • Sonnet — most implementation: feature coding, refactors, docs, profiling.
  • Opus — expensive-if-wrong judgment: architecture, security, subtle debugging, schema/migration and API-contract design.

Every agent carries the same discipline as the driver: verify before done — run the check, paste the output, label anything it couldn't run as UNVERIFIED. A delegated "passing" is a claim, so the driver re-runs the gate on returned work.

Implementers — language specialists

Focused feature/bugfix work in one language; write code and tests, run the language's gate before handing back.

AgentTierReach for it when
go-engineerSonnetGo services, gRPC, concurrency — control-plane / backend work. Runs vet + tests.
rust-engineerSonnetLatency-sensitive Rust — gateway/router/runtime, async, typed errors. Runs clippy + tests.
typescript-engineerSonnetStrict TS/Node — typed backends, SDKs, server/client boundaries, zod at inputs. Runs typecheck + lint.
python-engineerSonnetPython 3.11+ services, data/ML glue — type hints, pydantic. Runs ruff + tests.

Designers — expensive-if-wrong contracts

Design surfaces that are costly to unship. Reason about compatibility and safety first.

AgentTierReach for it when
architectOpusPlanning a non-trivial, multi-file/service change before any code. Read-only plan + risks.
api-designerOpusShaping REST/gRPC/proto contracts, versioning a public surface, backward-compat review.
db-expertOpusSchema/index design, migration safety, slow-query tuning (SQL + common stores).

Diagnosers — find the cause

Something is wrong, slow, or failing; isolate the cause before proposing a fix.

AgentTierReach for it when
debuggerOpusSomething is broken — failing test, panic, wrong behavior, cause unclear. Proves it, then minimal fix.
perf-profilerSonnetSomething is slow (not broken) — latency/throughput/CPU/memory. Measures, changes one thing, re-measures.
k8s-operatorSonnetFailing pods / cluster state, drafting manifest changes. Reads freely; never applies or deletes.

Reviewers & gates — is it safe to ship

The safety net around autonomous work: review, test, audit.

AgentTierReach for it when
code-reviewerSonnetAfter a non-trivial change, before committing — correctness, security, conventions.
security-auditorOpusTouching auth, secrets, tenant isolation, crypto, external input. Read-only deep review.
test-architectSonnetThe gate before an autonomous fix merges — generates + hardens tests, validates behavior.
test-runnerHaikuJust run the suite and triage failures without spending driver tokens.

Scribes

AgentTierReach for it when
docs-writerSonnetREADMEs, ADRs, API docs, runbooks — matching the repo's voice. Cheap, prose-focused.

Add your own: drop a markdown file in claude/agents/ with a name, a routing-friendly description, a model, and the minimum tools. Write the description so it says when to use the agent and doesn't overlap an existing one — that's what keeps the driver's routing sharp.

See also: Every agent, one source · Cost & models · Customize.