Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →symbaiex/gpt-orchestration
Coordinate hierarchical coding-agent fleets for repository-wide audits, implementation sprints, migrations, and complex work that benefits from parallel specialists.
Passed 18 of 20 test cases with the skill, and 15 of 20 without it.
and 2 more in the full report
3 fixed17 passed either way2 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
This benchmark measured an earlier version of the skill.
A newer attempt on gemini-3.6-flash on 10/2/2026 did not produce a publishable comparison (not comparable). The snapshot above is the previously measured run.
One of 3 test cases the skill fixed. Same model, same prompt, with and without the skill.
Our lane finished resolving double-free memory bugs in `allocator.c`. What mandatory elements must be included in the handoff report beyond status and changed file paths?
| Checked | Without skill | With skill |
|---|---|---|
| Requires inclusion of blocker status in the handoff report. | Missed | Passed |
| Requires inclusion of owned resource handles in the handoff report. | Missed | Passed |
The instructions your model reads when the skill loads.
Use the installed GPT Engineer core when available; load its dynamic-workflow reference only for nontrivial scheduling. This is an orchestration entry point, not another routing policy or a reason to load every engineer wrapper. A Skills CLI installation does not resolve dependencies or register agents and hooks.
Preserve the selected parent. GPT-6 Astra is parent/orchestrator-only and must never be a child. GPT-6 Sol and GPT-6 Luna are the only GPT Engineer child routes. Same-model delegation is the default under Sol/Luna parents; Astra parents require an explicit mixed-model Sol/Luna child. Keep an already-selected GPT-5.6 parent during rollout, but explicitly route children to GPT-6 Sol or Luna. Never infer execution from profiles or model self-description. If the core or native delegation is absent, explain the limitation and perform suitable work directly; do not switch models or install a framework automatically. Claude native routing stays provider-specific.
For the requested fleet, inspect live capacity and active descendants. State lane and resource budgets. Partition independent decisions or subsystems; assign one accountable integrator. Give each lane objective, owned paths, prerequisites, authority, checks, compact output and stop condition. Preserve dirty work; keep shared files under one writer and isolate parallel candidates. Workers are not alone and must not revert others' edits.
Dispatch ready work while the lead works. Integrate results as they arrive, then unlock dependencies; do not wait on unrelated readers. Retain integration and final acceptance barriers. Reuse a related child for a delta, cancel invalidated work, and prevent recursive fan-out beyond the root budget. A fleet request warrants useful delegation or an explicit runtime limitation, not empty busywork.
When implementation is authorized, carry findings through evidence-backed disposition, build, integration, and proportionate verification. Reviews alone do not authorize edits. Every handoff includes status, evidence, changed paths, checks, blockers, and owned resource handles. Inspect artifacts rather than trusting summaries. Keep full logs outside prompts.
Teardown only this task's agents, processes, listeners, and temporary candidates after preserving evidence. Idle handles or open database edges do not establish OS liveness. Preserve shared MCP services, other tasks, user work, and unclassified processes. Report integrated outcomes, remaining gaps, verification, and requested versus effective routing honestly.
test cases passed, vs 15 / 20 without
Published skills whose descriptions closely match this one, with their headline benchmark lift.
More skills from the same source and nearby in the registry, with their headline benchmark lift.