Skip to content

PATTERN Cited by 1 source

Agent orchestration meta-skill

Intent

Bundle the golden-path end-to-end workflow for a specific codebase into a skill — not a narrow single-tool binding, but a procedural runbook the agent consults for multi-step work on that codebase. The agent loads the skill on demand, not as part of a permanently-on system prompt, so the context cost is paid only when the skill is relevant.

Canonical articulation — Atlassian Fireworks, 2026-04-24:

"We've built a meta-workflow / orchestration skill for Fireworks development. It doesn't do one narrow technical thing, instead it gives the agent a set of 'golden path' loops for how to work on Fireworks changes end-to-end." (Source: sources/2026-04-24-atlassian-rovo-dev-driven-development)

Shape

   Agent conversation
        │
        │ "I'm making a Fireworks change"
        ▼
   Load fireworks-dev skill ◄──── addressable, not always-on
        │
        │ skill contents loaded as context:
        │   • entry workflows (make PR, fix test, deploy shard)
        │   • tool sequences for each workflow
        │   • verification checkpoints
        │   • common traps
        │   • escape hatches to human
        ▼
   Agent proceeds with loaded procedural context

What goes in the skill

Section Content
Entry workflows "Raise a PR", "Deploy to dev shard", "Debug a failing e2e test", …
Tool sequences Per entry-workflow: which tools to invoke in which order
Verification checkpoints When to pause and check (CI output, dev-shard health, tests)
Common traps / exclusions "Don't skip Helm validation", "Don't merge without e2e", …
Escape hatches When the agent should hand back to the human

The Fireworks team's second named orchestration skill illustrates the pattern at a narrower scope: "a skill that automates deploying, operating, and tearing down isolated Fireworks dev shards on the shared AWS scms Kubernetes cluster" — still multi-step, still procedural, scoped to the dev-shard lifecycle rather than the full codebase.

Not the same as

Narrow tool-binding skill Orchestration meta-skill
Scope One domain / one action End-to-end workflow across domains
E.g. "How to use our CLI", "How to write a Raft test" "How to work on Fireworks changes end-to-end"
Answers What does this tool do? Which tools in which order for this change shape?
Invocation Single use Multi-step loop with checkpoints

Why on-demand, not system-prompt

A codebase-specific meta-skill can be substantial (hundreds of lines of procedural instructions). If it were loaded into every conversation's system prompt:

  • Context budget inflates for every task, even ones that don't touch this codebase.
  • The agent has to attend to irrelevant instructions, reducing the signal-to-noise of its context.
  • Updating the skill requires rebuilding the agent's system prompt, which couples codebase evolution to agent configuration.

Addressable / on-demand skills avoid all three.

When it fits

  • Complex codebase with non-obvious golden path. The agent's default tool-sequence discovery is insufficient; pre-compiled procedural knowledge is worth the authoring cost.
  • Stable-ish workflow. If the golden path changes every week, maintaining the skill costs more than it saves. Orchestration skills pay off when the workflow is relatively stable even if the code underneath isn't.
  • Multi-agent team. Several developers running several agents on the same codebase — the skill amortises procedural onboarding across all of them.

When it doesn't fit

  • Simple codebase. If the golden path is "read, edit, test, commit", a skill is over-engineered. A README.md is enough.
  • Unstable workflow. Workflow changes faster than the skill can be maintained; the skill becomes a stale source of wrong procedural guidance.
  • Single-developer / short-lived project. Authoring cost is not amortised.

Relation to other skill patterns

Seen in

Last updated · 510 distilled / 1,221 read