What is this for? To explain when one agent should own the whole job vs when to split work across a multi-agent team.
Why does it exist? One giant agent with dozens of tools gets overloaded. Its context fills up, tool selection gets noisy, and tool hallucination (inventing or misusing tool calls) becomes more likely. Specialists are easier to trust than one overworked generalist.
One skilled generalist can close a ticket: research policy, draft a reply, file the note. That is single-agent.
A newsroom-style pipeline—researcher, writer, fact-checker—is multi-agent. The second can be higher quality at scale, but you now own handoffs, shared memory, and "who is stuck?" debugging.
| Plain-English idea | Single-agent | Multi-agent |
|---|---|---|
| Who does the work? | One loop, all tools | Several specialists, each with narrow tools |
| Tracing | One timeline | Must track which agent produced what |
| Default? | Yes—start here | Only when specialization clearly wins |
Default: start single-agent. Move to multi-agent when specialization and parallelism clearly beat the orchestration tax.
| Agent role | Plain-English idea | Typical tools |
|---|---|---|
| Triage agent | Front door that routes the task to the right specialist | Classifier, router |
| Infrastructure agent | Reads live telemetry, logs, cluster signals | Splunk, metrics APIs |
| Codebase agent | Checks commit diffs, pull requests, source code | GitHub, git |
| Policy & HR agent | Looks up runbooks, policy, knowledge base entries | RAG over docs |
Each specialist gets a small, relevant toolset—not every API in the company.
| Pattern | Plain-English idea | Watch-out |
|---|---|---|
| Pipeline | Fixed stage order | Brittle if a stage fails silently |
| Supervisor | Boss agent delegates | Boss becomes a bottleneck |
| Peer debate | Two agents critique | Cost explodes; needs a judge and stop rule |
| Router / triage | Classifier picks one specialist | Mis-routing causes confident wrong answers |
At enterprise scale, the triage agent becomes a meta-planner. Specialist agents become its "tools." The triage agent reasons at the strategy level and delegates the details—it does not call every API itself.
Stay single-agent while:
Consider multi-agent when:
Measure before splitting: if a single agent plus a deterministic reviewer script hits your bar, skip the second LLM.
A supervisor that delegates to two specialists with a shared artifact—still simple enough to debug.
from dataclasses import dataclass
@dataclass
class Artifact:
brief: str = ""
notes: str = ""
draft: str = ""
verdict: str = ""
def researcher(art: Artifact) -> Artifact:
art.notes = f"facts for: {art.brief}"
return art
def writer(art: Artifact) -> Artifact:
art.draft = f"Draft based on [{art.notes}]"
return art
def reviewer(art: Artifact) -> Artifact:
art.verdict = "pass" if "facts" in art.notes and art.draft else "fail"
return art
def single_agent(brief: str) -> Artifact:
art = Artifact(brief=brief)
return reviewer(writer(researcher(art)))
def multi_pipeline(brief: str) -> Artifact:
art = Artifact(brief=brief)
for stage in (researcher, writer, reviewer):
art = stage(art)
if art.verdict == "fail":
break
return art
print(single_agent("weekly support themes").verdict)
print(multi_pipeline("weekly support themes").verdict)Frameworks differ in APIs; the invariant is explicit artifacts and stop conditions.
Run an honest A/B on your golden set: single-agent baseline vs a two-role pipeline (implementer + reviewer). Compare pass rate, p95 latency, and cost per success. Promote multi-agent only if the quality lift beats the cost/latency hit.
If you do split, publish a one-page orchestration diagram: roles, tools per role, artifact schema, and max turns. On-call engineers should debug from that page at 2 a.m.
Rough planning math helps. If each agent turn costs about C dollars and takes T seconds, a supervisor plus three specialists that each speak twice costs roughly 8CT in the worst chatter pattern. A single agent that makes four tool calls costs about 4CT plus tool time. Unless eval quality rises enough to justify the multiplier, the multi-agent bill is vanity. Write the inequality down before the rewrite.
Use one agent until specialization clearly improves measured quality or safety; only then add multi-agent roles with typed handoffs, narrow tools, and hard stop rules.