Identify the multi-agent topology most prone to deadlock and the minimal defence

Attempt it

**Deadlock** is when two or more agents are each waiting for the other to produce a result, so nobody makes progress. It is different from a talk loop (where agents keep talking but go nowhere) and different from exceeding max_turns (which counts actual turns). The topology that allows deadlock is the **peer swarm**. Each agent owns its own routing decisions; there is no supervisor watching the whole graph. So if agent A's next step is 'wait for agent B's answer' and agent B's next step is 'wait for agent A's answer', nothing happens. Both agents are technically running; they are just both waiting. Tree topologies (supervisor + workers) almost never deadlock because the supervisor always picks the next worker. Sequential processes (one task at a time) and round-robin (fixed order) cannot deadlock for the same reason: someone or something always has the next turn. The minimal defence is a **watchdog timer**. Track when each agent last did real work. If an agent has been idle (waiting, not producing) past a threshold, fire a deadlock-break handler. The handler can return an error, force a handoff to a designated arbiter agent, or terminate the run cleanly with a message to the user. You need this on top of max_turns, not instead of it.

**Peer swarms** are the topology where deadlock is a real risk. In a peer swarm, agents transfer control to each other directly (handoff tool calls in OpenAI Agents SDK, swarm prebuilts in LangGraph) with no central coordinator. If agent A's continuation requires B's result, and B's continuation requires A's result, neither makes progress and the run hangs. The other options each have a built-in deadlock preventer. A supervisor topology has the supervisor as the always-ready next-decider; even if a worker is slow, the supervisor's loop is alive. Sequential CrewAI processes are strictly linear by construction. Round-robin enforces a fixed rotation, so the next agent is determined before any wait happens. The minimal defence is a **watchdog**: per-agent inactivity timer that fires a break handler once any agent has been waiting past a threshold (5 to 30 seconds in practice). The handler typically does one of three things: terminate the run with an error, force a handoff to a designated arbiter agent that decides who continues, or roll back to the last checkpoint and retry with a slightly different prompt. Watchdog complements `max_turns`; it does not replace it. `max_turns` catches runaway loops where turns are being taken; the watchdog catches deadlocks where no turn is being taken at all.

Quick reference

TL;DR

Peer swarms deadlock when agents wait on each other with no central coordinator. The minimal defence is a watchdog timer on per-agent inactivity, not just max_turns.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Two kids at a doorway, each politely insisting the other goes first. Neither moves. Nobody is being rude, nobody is stuck in an infinite loop; they are just both waiting. That is what deadlock looks like in a peer swarm of agents. A supervisor topology avoids it because a parent figure is always making the next call. The fix in the swarm is a hall monitor with a stopwatch: if nobody has moved for 30 seconds, the monitor steps in, picks one kid, and tells them to go. The stopwatch is the watchdog, and you need it as well as a hard time limit on the whole game, because the game can be inside its time limit while the kids are still frozen at the door.

Key concepts

Multi-Agent Systems

Detailed answer & concept explanation

~5 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

6 to 8 min: why peer swarms deadlock and others do not + the difference from loops and from max_turns overruns + watchdog mechanics + the arbiter handler + complementary defences (cycle detection, typed protocols, timeouts).

Real-world examples

Real products, models, and research that use this idea.

OpenAI Agents SDK swarm-mode deployments routinely add a watchdog wrapper that catches inactive-agent states and forces a handoff to a designated arbiter.
LangGraph swarm tutorials warn explicitly about the cycle of handoffs failure and recommend a max_handoffs counter plus inactivity timer.
Devin and Cursor agent traces show watchdog-fired terminations as a tracked metric, not as a rare event.
Replit Agent's 2025 incident reports cited peer-swarm deadlock as the failure that drove their adoption of supervisor plus workers as the default topology.
Microsoft AutoGen 0.4's termination condition primitives include time-based gates partly to defend against the same class of failures.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you build an arbiter agent that handles deadlock without making the system worse?

Sign in or go Pro to see the approach for this follow-up.

QHow does deadlock risk change when you move from in-process agents to distributed agents over a message bus?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Assuming max_turns prevents deadlock. Max_turns counts turns taken; if no agent is taking a turn (both blocked on each other's result), the counter never advances and the run hangs until the wall-clock kills it.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.