kinds of agents

coordinator

• never delegates understanding
• the coordinator must synthesize research into specific specs with file paths and line numbers.
• the phrase, "based on your findings, fix it" is explicitly banned in the system prompt.
• the smartest available LLM should implement this role
• must use reasoning/thinking

researcher

• does not have write permissions; cannot do any actual work
• can search the web, can read filesystem, can use limited commands like git
• use a faster LLM with less intelligence
• should use reasoning/thinking

implementer

• can change files
• this agent does all of the work assigned to it by coordinator
• does not require reasoning

verifier

• fresh agent
• the smartest available LLM should implement this role
• must use reasoning/thinking
• should be skeptical

orchestration pipeline

1. research: subagents operate in parallel
2. synthesis: coordinator evaluates responses from research subagents
3. implement: subagents operate serially, per file set
4. verify: a separate agent (not the coordinator) compares the implementation to the spec

scratchpad

• all agents may read and write to var/plans
• enables durable cross-worker knowledge sharing