---

name: ztx5-multiagent-agent-cost-control description: "Multi-agent concurrency/cost. Use when building multiagent systems or when an audit detects this specific problem."

Multi-agent concurrency/cost

Trigger: Whenever the product calls for multi-agent concurrency/cost.

Problem and behavior contract: Run only independent agents in parallel; cap workers and duration; exit with errors clearly.

Implementation: ThreadPoolExecutor(max_workers=4) and subprocess timeouts, no nested recursion.

Execution procedure

1. Inspect the existing implementation, UI state ownership and dependencies; identify exact target component and current behavior.
2. Read ../../references/MULTIAGENT-ENGINEERING.md for compatible APIs and full code examples (when installed, use ../../ztx5-resources/references/ instead) (when installed, use ../../ztx5-resources/references/ instead) (when installed, use ../../ztx5-resources/references/ instead), and consult linked original official sources in SOURCES.md.
3. Produce a small technical plan containing target elements, input events, animation timebase, responsive formula, cleanup, and fallback.
4. Implement only the relevant components and avoid global DOM/CSS overrides; keep a route unmount/teardown function.
5. Capture normal, reduced-motion, mobile, desktop, RTL and LTR screenshots when a renderer is available.
6. Document test observations, unresolved dependency requirements and what changed.

Specific acceptance

Timeout returns partial report and nonzero code.

Deliverables

• Actual working project code (not merely a paragraph)
• Local verification script or Playwright check
• At least one computed screenshot or recorded property check
• PASS/FAIL/NOT_TESTED status with reasons

Upstream research

• https://github.com/frankxai/awesome-design-agent-skills
• https://github.com/hueyexe/frontend-agent-skills
• https://developers.openai.com/api/docs/guides/tools-skills

Hard constraints

• Every skill must preserve previous project behavior. No deleting or wholesale replacement.
• Actual user-provided brand rules outrank any stylistic recommendation. If absent, request identity; no fake approvals.
• Responsiveness: test 320, 360, 390, 430, 768, 1024, 1440, 1920 CSS px; RTL and LTR; 200/400% zoom when feasible.
• Reduced motion, touch, keyboard and failure/no-JavaScript fallback must be considered.
• Performance: avoid forced layout in rAF; disable noncritical GPU motion on weak devices.
• Choose one timeline/transform owner; GSAP, Anime.js, Motion and native CSS must not race on same property.

Do not mark complete unless

1. Actual integration code is committed into the target project, or clearly labeled as snippet-only.
2. A build/compile test passes in the installed runtime.
3. Browser output is inspected in 2 languages and at least 2 viewport classes.
4. Acceptance tests pass with explicit evidence files; report incomplete tests honestly.