Mihira Jayasekera@mihira_j
Agent-builder; ex-researcher
Multi-agent research workflowsPlanner + specialists
Representative of practitioner threads on running a planner agent that delegates subtasks to reader, writer, and verifier agents — each with a narrow prompt, all coordinated by a shared scratchpad.
“One giant prompt is a junior analyst. Four small agents are a team.”
How to replicate
- 1.Define 3–4 narrow roles (planner, reader, writer, verifier) with a one-paragraph system prompt each.
- 2.Use a shared scratchpad (JSON file or KV) as the handoff mechanism.
Prompt template
You are the <role> in a 4-agent research team. Your job: <one paragraph>. You may read from scratchpad keys: <list>. You may write to key: <one key>. Do not do the work of other roles. If you cannot proceed, write to `blocked` with a reason. Current state: <paste scratchpad>.
Pitfall
Letting every agent write to every key. Without role-locked keys, specialists overwrite each other's work and the planner loses track of state.
What you'll learn
- •Why narrow roles beat one omniscient prompt
- •How a scratchpad replaces long message histories
- •When to add a verifier role vs relying on self-critique
- •How to cap multi-agent loops before they cost you hundreds of dollars
#multi-agent#research#orchestration