𝗔 𝗙𝗶𝗲𝗹𝗱 𝗚𝘂𝗶𝗱𝗲 𝘁𝗼 𝗠𝘂𝗹𝘁𝗶 𝗔𝗴𝗲𝗻𝘁 𝗙𝗮𝗶𝗹𝘂𝗿𝗲 𝗠𝗼𝗱𝗲𝘀

𝗔 𝗙𝗶𝗲𝗹𝗱 𝗚𝘂𝗶𝗱𝗲 𝘁𝗼 𝗠𝘂𝗹𝘁𝗶-𝗔𝗴𝗲𝗻𝘁 𝗙𝗮𝗶𝗹𝘂𝗿𝗲 𝗠𝗼𝗱𝗲𝘀

Stop saying agents got confused. Vague words do not help you fix things. You need a clear list of failures.

Cemri et al. studied 1,642 traces from 7 frameworks. They found 14 failure modes in 3 groups.

Group 1: Specification failures. These happen at design time. The agent follows a bad setup.

• Step repetition (15.7%)
• Ignoring tasks (11.8%)
• Ignoring roles
• Losing history
• Missing stop rules

Fix these first. They are cheap.

Group 2: Coordination failures. These only happen with multiple agents.

• Agents ignore each other.
• Agents hide info.
• Reasoning and action mismatch (13.2%).

Fix: Share full execution traces. Do not share only messages.

Group 3: Verification failures. These have high impact.

• Early stops.
• No checks.
• Wrong checks.

Fix: Add a verification step. This raised success by 15.6%.

Finding the bug is hard. Zhang et al. found low accuracy for failure attribution. Agent accuracy was 53.5%. Step accuracy was 14.2%.

Mistakes cascade. One early error causes a later crash.

Your plan:

• Fix specs first.
• Add a verification step.
• Fix coordination with architecture.

Source: https://dev.to/tuomo_pisama/a-field-guide-to-multi-agent-failure-modes-59on