- Home
- Skills
- Agents & Orchestration
- agentic-maestro
More screenshots
Works with the AI Agents you already use
Agentic Maestro
Delegate coding work to cheaper sub-agents without losing quality: precise briefs, independent review, one launcher for Claude Code, Codex and OpenCode.
Free
agentic-maestro
Example session with this skill installed
Use agentic-maestro to add page-size limits to our export API. Delegate the build to a cheap executor on Codex, have Claude review it, and merge only if it passes.
- Read your context and instructions
- Compiled the agentic-maestro
Plan — 1 lane, delegated (touches 3 files and needs tests, so a brief beats doing it inline)
| Ruling | Exact semantics |
|---|---|
| R1 | pageSize is an integer 1..500; anything else is refused with EXPORT_PAGE_TOO_LARGE (400), never clamped |
| R2 | The bound is checked before the database query runs |
| Contract | exportRows(query) keeps its signature; existing error codes are unchanged |
Done means: 500 succeeds; 0 and 501 are refused; no query is issued for a refused request (counted with the query spy).
Brief assembled and checked
$ node scripts/brief.mjs F12.json --out .agent-briefs/F12-build-r1.md
brief.mjs: wrote .agent-briefs/F12-build-r1.md (5.2 KB)
It carries the shared rules, the two quoted conventions that apply, the Input and Tests checklist areas, one pattern excerpt (packages/import/src/limits.ts), the 3-file map and the per-package gates.
Executor (Codex, cheap model, own worktree)
$ node scripts/spawn.mjs --harness codex --effort high --brief .agent-briefs/F12-build-r1.md --dir ../wt-F12 --out .agent-runs/F12-build-r1
spawn.mjs: codex ok exit=0 41s session=01a1539a-…
Report block
STATUS: done
FILES: packages/export/src/limits.ts, packages/export/src/handler.ts, packages/export/test/limits.test.ts
R1: done
R2: done
GATE pnpm --filter @app/export test: exit=0 passed=48 failed=0
GATE pnpm --filter @app/export typecheck: exit=0
CHECKLIST: ok
BLOCKED: none
Scope check before review: scope-check.sh ../wt-F12 3f2c1ab 'packages/export/' → exit 0, nothing outside the map. Committed as 9c41e07.
Independent review (Claude, read-only) on 9c41e07: 1 finding.
- P2 —
limits.test.ts:31tests 501 but not 0 →
AGREE. Fix ruling: add the lower-bound test asserting EXPORT_PAGE_TOO_LARGE and zero queries.
Fix round 1 (fresh executor, fix brief with the finding quoted): 1 test added, gates exit 0.
Re-review in the reviewer's own session: PASS.
Merged 9c41e07..b18d2aa into main. Ledger: 2 executor runs, 1 review + 1 re-review, 0 blocked questions.
Next steps: regenerate docs/CODEMAP.md (a file was added), and add the clamping finding's category (input/clamp) to the project checklist if it shows up in a second feature.
Connects securely to your tools. The creator never sees your data.
About this skill
Your strongest model shouldn't write every line. agentic-maestro turns it into a conductor: it plans, writes exact rulings, arbitrates review findings and merges, while cheaper sub-agents execute precise briefs and an independent reviewer checks every result.
What you get:
- When to spawn at all. A rule for when delegation pays and when the orchestrator should just do it inline.
- Briefs instead of context. Templates for build, fix and integration briefs: exact rulings, a "done means" list, one pattern to copy, only the conventions that apply, and a fixed report block you can scan in seconds.
- Fail-closed tooling.
brief.mjsrefuses a brief with a missing value or a convention keyword that matches nothing;scope-check.shflags files changed outside the brief's map, even ones hidden by renames or git index flags. - One launcher for every agent.
spawn.mjsruns Claude Code, Codex or OpenCode (v1 and v2) headless in write or read-only mode, resumes exact sessions, and records session id, exit code, tokens and cost per run. - A pipeline that converges. Independent review from a different model family, arbitration of every finding, a fresh executor per fix round, a structural ruling after three rounds on the same component, and a checklist that learns from recurring findings.
- Parallel lanes on one machine. File ownership per lane, a lock for heavy checks, stacked branches and small merge batches.
Measured on one large TypeScript monorepo (field notes, not a benchmark): 65–70% fewer tokens re-read by executors, about half the build time per feature, and fix rounds down from ~74 min to 10–25 min. Lean headless mode cut a one-line Claude Code run from $0.15 to $0.06.
Works with Claude Code, Codex, OpenCode and any Agent Skills host. Requires Node.js 18+ and bash. Source and docs: https://github.com/Akramovic1/agentic-maestro
Needs git and Node.js 18+; give each writing executor its own git worktree. Semantic merge conflicts between lanes are reported to the orchestrator, never auto-resolved. The code map lists exports only for JS/TS workspaces (other languages get a file-level map). The measured gains are field notes from one TypeScript monorepo, not a benchmark.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 28 days ago
- Free with an Agensi account
Requires
Permissions
Terminal: the scripts run git, node and the agent CLI you choose (claude, codex or opencode) in a working directory you pass. Read/Write files: briefs, run logs and reports are written inside your project (.agent-briefs, .agent-runs, .agent-logs), and executors edit only the files in their brief's map, ideally in a disposable git worktree. Network: the scripts make no network calls of their own; the agent CLIs they start talk to their own model providers with your existing login. No secrets read, nothing installed.