One agent from triage to pull request cut Polylane's median detect-to-PR time to 35 minutes

Follow one incident through the March design. A triage agent decides whether the issue is real. A coordinator writes candidate explanations. Up to 15 sub-agents test those hypotheses from different starting points, confidence-weighted voting combines what they found, and the result is summarized for the coordinator ^[4]. If a likely root cause emerges, a coding agent receives a plan and attempts the fix ^[4]. It does not necessarily receive the original issue or the evidence that produced the plan ^[5].

Polylane built that pipeline in March 2026, when frontier models struggled to carry an infrastructure problem from investigation through a code fix and its own agent harness was less mature ^[9]. Splitting the job gave each agent a narrower one. The cost showed up in the joins. The Neuron's account notes that a summary keeps the conclusion and loses the observations a later agent would need to argue with it ^[18]. Polylane says agents regularly repeated work and the final coding agent produced subpar pull requests that often attacked symptoms instead of root causes ^[6].

Change the denominator from pull requests to detected issues and the cost picture flattens. Before the cutover, 0.6 percent of detected issues reached a pull request at $111 of model spend each, about 67 cents per detected issue ^[13][14][26]. After, 4.2 percent reach one at about $18, about 76 cents ^[26]. Per-PR spend fell about 84 percent while spend per detected issue rose about 14 percent ^[27][26]. Both halves have to come from their stated comparison windows for that to hold, and the $18 covers the first nine days after the September 3 cutover ^[14].

Polylane says other product changes contributed to the lower costs, so the redesign cannot claim the whole reduction ^[15]. The pull-request rate counts proposals, and it does not say how many were accepted or resolved the underlying issue ^[16].

For the speed numbers to transfer, your later steps have to depend on evidence your earlier steps gathered. The Neuron's second account draws the line at exactly that dependency: separate research searches or independent code reviews split across workers more safely, while Polylane's infrastructure investigations depend on evidence collected early ^[19]. Anthropic still recommends Claude Code subagents for tasks that would otherwise fill the main conversation with search results or logs, each subagent getting its own context window, custom instructions and potentially restricted tools ^[20]. LangChain's guidance says one agent can handle many complex tasks, and keeps the subagent pattern for cases where a supervisor combines independent results afterwards ^[21]. The rule in the first account is that extra agents are worth adding when they add genuinely independent work, evidence or judgment ^[23].

The part I would carry into any topology is the evaluation problem. Individual stages passed their own evaluations while the overall result stayed weak, because the failures lived in the handoffs ^[7]. Engineers sometimes had to inspect dozens of agent traces to understand why one investigation went wrong ^[8]. Polylane's own explanation for the improvement is that the agent making the final decision now retains the investigation that produced it ^[24].