Why does a repair loop need a spend ceiling and not only an attempt limit?

Because attempts are not equal in cost. Salesforce stores budgets, turn limits, and cost ceilings per tier and construct type in one configuration file and applies a topology-specific formula, a flat check for one agent, items times cost times steps for a swarm, a whole-roster ceiling for a team, so a repair cannot commit to a plan it cannot afford to finish. The $3 ceiling sits on top of the two-attempt limit for that reason.

What does LangSmith Engine v2's repair loop do, and what is missing from the announcement?

For agents on LangSmith Deployment it reruns the offending inputs to confirm an issue, tests candidate fixes against a broader eval set until it finds one that resolves it, and presents the fix for a one-click PR. Jacob Talbot's post states no maximum number of candidate fixes, no spend ceiling, and no named escalation. It may exist in the product; it is not in the post a buyer reads.

Why is the stop rule a product decision rather than an engineering one?

Because the test that decides whether a repair worked compares output to expected behavior, and expected behavior for the failures that matter, Salesforce's soft failures that execute cleanly and return the wrong result, is a bottom-up eval a human writes by reading real outputs. Who writes that file, and when the loop must hand a case back to a person, are decisions about the product, and they are made before the loop runs or not at all.

What did the pricing agent without a stop rule cost?

Five weeks of compounding optimization toward a local maximum left the price 14% below what the benchmark said was correct, and recovering the anchor took two quarters of careful re-pricing. Every iteration passed its own test. The rule since is a human approval gate at every iteration for autonomous optimization in long-feedback-loop domains.

What Stop Rule Should a Self-Healing AI Agent Have

What stop rule should a self-healing AI agent have?

THE SHORT ANSWER

Three fields, decided by the product owner before the loop is switched on: the maximum number of repair attempts, the spend ceiling per incident, and the name of the person the failing case escalates to with the diagnosis attached. Salesforce's Agent Designer, per Sohini Arya and Manish Kumar Jha, limits self-healing to two fix attempts and a $3 ceiling inside a 500-turn orchestrator budget, then stops and escalates. Underneath the three fields sits a domain rule: anything touching money, contracts, individual customers, or decisions that compound gets a human review layer that configuration cannot remove. My autonomous pricing agent had none of this, drifted 14% under benchmark in five weeks, and took two quarters to recover.

Two self-repairing agent systems shipped in the same week, and The Stop Rule is about the one line that separates them: Salesforce's Agent Designer publishes two attempts, $3, and escalate, and LangSmith's Engine v2 publishes a loop that runs until a fix passes. The bill for running the second kind is agent number four in 10 AI Agents I Built That Failed. The Honest Retrospective., and the reason the expected behavior a repair loop tests against has to be written by a person is in The Wince Is the Spec: Bottom-Up Evals a Model Can't Write.

It belongs to the Enterprise AI Agents argument: what matters is how many deployed agents still complete production work after ninety days, and what each outcome costs. A loop with no stop rule has no cost per outcome.

Last reviewed 2026-09-29 · 1 min read

What stop rule should a self-healing AI agent have?

SOURCES

THE LONG VERSION

RELATED ANSWERS