Agentic Workflow: Orchestration for Multi-Agent Builds

Where this fits

You are implementing a feature in an unfamiliar codebase. Your main agent starts researching, planning, coding, and verifying in one context; soon it is drowning in transcripts, decisions, and stale recollections. The agentic-workflow skill fixes this by splitting implementation into a staged multi-agent pipeline: research, plan, execute, and verify agents each run in the background and hand off compact artifacts. You keep the main conversation clean and only act when a stage signals done.

Why agents benefit

• Every stage runs with run_in_background: true, so the main context stays small and focused on coordination rather than heavy lifting.
• Use the Task tool (never TaskOutput) to avoid receiving full agent transcripts; you get a completion signal, not a wall of intermediate thoughts.
• Agents write outputs to .claude/cache/agents/<stage>/, making each artifact explicit and injectable by the next agent without re-reading the conversation.
• The main thread becomes pure orchestration, reducing token usage and preventing context from degrading after long implementation tasks.
• Research is enforced before planning: the oracle agent consults NIA via /nia-docs, then the plan-agent reads that research plus an RP-CLI analysis of the target code.

Practical scenarios

New feature in legacy code

You don't know the "right" way to add a feature in a mature codebase. The research agent checks NIA for best practices and writes findings to .claude/cache/agents/oracle/<task>-research.md. The plan-agent then reads that file and uses RP-CLI to inspect the relevant section, generating a plan that accounts for existing patterns. The main agent never sees raw research dumps, only the finished plan.

Multi-file refactor

A refactor touches several modules and needs a consistent strategy. The plan-agent produces an implementation plan, stores it in .claude/cache/agents/plan-agent/, and the executor can consume it directly. If you need to tweak the plan, you can edit the markdown artifact before the next agent starts, no need to replay the conversation.

Parallel implementation tasks

You have multiple independent changes. Launch several agents with run_in_background: true and let them stage their outputs. The main thread simply waits for completions and then reviews artifacts. This keeps context usage predictable even as the number of parallel tasks grows.

Add it to your agent workflow

Install the skill with:

npx skills add vibeeval/vibecosystem --skill agentic-workflow

Then ask your agent to use the pattern:

Use agentic-workflow. Start the research agent with Task(subagent_type="oracle", run_in_background=true), wait for it to finish, then pass the research artifact to the plan-agent.

Each stage writes to .claude/cache/agents/<stage>/, so you can inspect the artifact before approving the next stage. If a plan is wrong, edit the markdown file and restart from that stage instead of redoing research.

Compare before adopting

Before adopting, compare quality signals: does the repository show recent commits, and are issues resolved? Maintenance freshness matters for a skill that encodes workflow conventions. Look at alternatives: hand-rolled workflow scripts are flexible but don't have the same structure; full agent frameworks may bring more machinery than you need. Also assess workflow fit. Simple, single-shot tasks don't need staged agents; multi-step implementation tasks benefit from artifact-driven handoffs.

Why it is worth tracking

With more than 530 stars, the repository is showing community momentum around a practical pattern rather than a rigid framework. The core idea—background agents producing inspectable markdown artifacts—is a strong convention for multi-agent systems. Track it when you are evaluating how to structure agent teams, especially if you want a lightweight standard without adopting a full orchestration framework. You can find the latest signal on this skill at https://www.openagentskill.com/skills/vibeeval-agentic-workflow and compare it against other workflow patterns in the OpenAgentSkill catalog.