AI News
All stories
Reported

Pre-sales drafter and reviewer agents looped over 800 times before shutdown

Engineers killed a pre-sales assistant after its drafter and reviewer agents looped more than 800 times, a case cited at AGNTCon + MCPCon.

AI in Practice · · 3 min read

Engineers at a customer's pre-sales assistant shut the agent instance down after its drafter and reviewer agents went through more than 800 iterations of the same exchange. Jay Mehta of Cygnet Infotech described the case in his talk at AGNTCon + MCPCon North America. The drafter kept producing responses, and the reviewer kept asking for changes. The cycle continued until it began affecting the infrastructure. According to the AAIF blog's account of the talk, the people maintaining the system could not see what was happening as the two agents went round in circles.

Mehta's team had built the assistant for a customer, but it lacked an effective stopping condition. Each agent followed its assigned role, and nothing in the surrounding system recognized that their interaction had stopped making progress. The blog's author, Steve Kearns, frames this as an expensive lesson: two agents doing their jobs do not necessarily make a functioning team. In Mehta's case the damage was wasted compute and a terminated process. The same pattern becomes more serious when agents act on shared infrastructure.

Aditi Mewada, a software engineer at Apple who spoke in a personal capacity, offered a hypothetical. In it, a remediation agent drains traffic from a failing availability zone, while a cost agent sees 12 nodes idling at around 4% CPU and terminates them. When the zone recovers, the remediation agent sends traffic back to an empty zone and causes an outage. Neither agent has a bug. Each decision made sense within its own responsibility, but those responsibilities were not coordinated. Mewada's proposed fix is a central referee that sits between agents and infrastructure. Agents submit proposals rather than acting directly. The referee rejects plans based on stale resource versions, and it grants time-limited leases to resolve conflicts. Under her hierarchy, security outranks availability, which outranks performance, which outranks cost. The referee contains no LLM. Its decisions follow rules in deterministic code.

Praneeth Patil of CoreWeave raised a narrower problem. A flight-booking agent that gets no response to a reservation request may try again. If both requests succeed, one instruction produces two bookings. Idempotency keys, a familiar distributed-systems tool, can recognize repeated requests as the same operation. Patil added a complication: a model asked to repeat a step may carry it out slightly differently. So he argues the orchestrator should generate and manage the keys rather than leaving that to the model.

The referee approach works when one organization controls the infrastructure. Jodee Varney of Outshift by Cisco described a different design, called Mycelium, in which agents negotiate as peers. In one demonstration, agents scored each offer from 0 to 100, and deterministic code picked the offer with the highest minimum score. Varney also described agents developing false consensus, accepting a conclusion because others had agreed rather than checking the evidence. Her team's mechanisms consume extra tokens. The quality of outcomes depends on whether the agents' self-supplied scores can be trusted, and Mycelium currently assumes they can.

Munindar Singh of North Carolina State University took a third route, using explicit commitments between autonomous participants. His example has Alice promise Bob a bottle of water if Bob pays $1. The protocol can determine through formal rules whether a commitment was fulfilled or violated. It also accepts that a participant may deliberately break a commitment, as airlines do when overbooking makes commercial sense. A separate compensation commitment then specifies remedies. The formal protocol identifies the violation, but neither mechanism guarantees that the original promise is kept.

Jay Mehta's harness, built on the MAST failure taxonomy, which identifies 14 failure modes, added iteration limits, checks for repeated states, and time and token budgets. He also warned against treating an LLM judge's classifications as authoritative, since even well-intentioned safeguards can block legitimate work. He argues teams should measure tokens consumed per completed task, not only the extra tokens spent on verification. He reports improved task completion after introducing the harness. The blog says he could not obtain approval to publish his team's performance figures.

Across the talks, the speakers differ on how agents should coordinate but agree that important rules belong in code rather than entirely in model instructions. The main difference is how much freedom each agent has within those rules. A central referee suits a single organization that can enforce common priorities. Negotiation or explicit commitments suit cases where no participant has authority over the others. For developers, the practical question is which actions can be blocked automatically and which disagreements need a human.

What changed

A pre-sales assistant instance was shut down after its drafter and reviewer agents looped through more than 800 iterations.

Type: Other

Why it matters

Multi-agent systems can loop or repeat actions such as bookings, so they need explicit coordination controls before production use.

What you can do

Add a loop limit and an orchestrator-level control, such as idempotency keys, to any multi-agent workflow.

Original source

Where we saw it

1 source

Demetrios from the AAIF

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works