Role specialization — planner, executor, critic, verifier
Planner
reads goal
produces spec
(signature, tests, I/O)
MetaGPT: PM + Architect
Executor
reads one step
produces artifact
has the real tools
ChatDev: asks if unclear
Critic
subjective review
catches taste issues
fooled by plausible prose
(LLM-based)
Verifier
deterministic check
tests, type, schema
cannot be fooled
(code, not LLM)
communicative dehallucination (ChatDev)
why verifier is load-bearing
MAST taxonomy (Cemri et al. 2025): 21.3% of 1642 multi-agent failures are verification gaps
PwC 2025: structured validation loop in CrewAI moved accuracy 10% -> 70% (7× gain)
anti-pattern: every role is an LLM that says "looks good to me" — classic MAST failure
always pair critic (subjective) with verifier (deterministic); they catch different bugs
MetaGPT's SOP pattern
Code = SOP(Team): Product Manager, Architect, Project Manager, Engineer, QA Engineer
each role has strict input/output schema · pipeline becomes deterministic
CrewAI Agent(role, goal, backstory) is the modern ergonomic surface
LangGraph encodes roles as graph nodes with specialized prompts