AI & agents02 / 12
Agent Studio
A Claude Code engineering studio with deterministic routing and hard gates
- Status
- Working, not deployed
- Period
- 2026 →
- Role
- Design and development, end to end
- For
- Internal tooling for my own projects.
Experience it
Illustration of real tooling: a scripted walk-through, not a recording of a live session.
One task through the studio
Current step · Step 1 of 16
Request arrives
A task is typed into the session: add a small feature to an existing project.
Message · You → Main agent
Step log
- 01 Request arrives[Message]
- 02 SessionStart injects the project briefing[Hook · allow]
- 03 Main agent plans[Message]
- 04 Delegates a read-only search to Explore[Delegate]
- 05 Explore reports back[Message]
- 06 Delegates the change to implementer[Delegate]
- 07 Implementer tries a shell command[Tool call]
- 08 guard_bash asks or denies[Hook · blocked]
- 09 Implementer creates a new file[Tool call]
- 10 guard_reuse blocks the duplicate[Hook · blocked]
- 11 Implementer extends the existing file[Tool call]
- 12 Tests run[Tool call]
- 13 Stop gate fails[Gate · fail]
- 14 Agent fixes the cause[Tool call]
- 15 Stop gate passes[Gate · pass]
- 16 Done[Message]
Problem
Working with AI agents is unreliable when every session improvises its own model choice, rules and checks. Quality depends on luck and on how carefully the last prompt was written.
Agent Studio is the working environment I build with: three plugins, thirteen specialist agents and sixteen skills, wired together with hooks that stop bad changes before they land.
Case study
Each agent pins its own model and effort level in its definition, so routing is a fact of the setup and not a decision made per session. An architect runs on Opus at extra-high effort, an implementer on Sonnet at medium, a scribe on Haiku at low, and a separate escalation tier exists for the hard cases.
Hooks enforce the rules. They deny dangerous shell commands and enforce conventional commits, keep secrets and lockfiles out of reach, block duplicate files and definitions on the first attempt, and run the project's own lint, validation and test gates before a turn may end.
Architecture
Three plugins (studio, studio-web, studio-ai) hold the agents and skills. Seven hook registrations across five events implement the guards for bash, files, reuse and the stop gate, which allows at most three attempts and scans for placeholders.
Engineering decisions
- Model and effort pinned per agent, so cost and quality are predictable.
- Rules enforced by hooks that act, not by instructions the model may forget.
- A reuse guard that blocks duplicates on the first attempt.
- The writer and the reviewer are always different contexts.
Challenges & trade-offs
Making guards strict enough to matter without making them a source of friction, which is why the stop gate caps retries and the hooks are covered by their own tests.
Verified results
- 3 plugins, 13 specialist agents, 16 skills
- 7 hook registrations over 5 events, 31 hook tests
- Working locally; private
Status labels are factual: nothing here is presented as more finished than it is.