Designing Reliable Multi-Agent Systems

Sample article. This is illustrative placeholder content. It is not company research or a report of customer experience.

Splitting a task across several agents can make each part simpler to build and test. It also adds coordination problems. Reliability comes from treating the system as a graph with explicit contracts, not as a conversation.

Start with one agent

Add a second agent only when a clear boundary justifies it: different tools, different risk levels, or a prompt that has grown too large to evaluate. Otherwise the extra hand-offs just add failure modes.

Give every agent a contract

Each agent should declare its inputs, outputs, permitted tools, and budget.

from typing import TypedDict

class ReportState(TypedDict):
    question: str
    sources: list[dict]      # written by retriever
    findings: list[str]      # written by analyst
    review_notes: list[str]  # written by reviewer
    status: str              # "drafting" | "needs_human" | "approved"

Share typed state, not free text

Passing structured state between steps makes each hand-off inspectable and testable. Free-text relays tend to lose information and make failures hard to localise.

Bound everything

• Loops: cap iterations between agents such as writer and reviewer
• Budget: limit tokens and tool calls per run
• Time: set deadlines and decide what happens when they pass

Make hand-offs traceable

Record which agent acted, what it saw, what it decided, and why. A trace that spans the whole graph lets you find the step where quality dropped.

Failure mode	Typical cause	Mitigation
Endless back-and-forth	No iteration limit	Max rounds, then escalate
Lost context	Free-text hand-offs	Typed shared state
Over-permissioned agent	One shared credential	Per-agent tool scopes
Silent quality drop	No evaluation	Regression suite on the full graph

Keep a human exit

Give the supervisor an explicit “needs human” outcome and make sure the state needed to resume is checkpointed.

Related reading: Building Production-Ready AI Agents and AI Agents vs Traditional Automation.