Sub-Agents Are Just Wrong

Polylane describes building an automated “autofix” system to scan logs, metrics and traces across cloud resources and resolve deviations from historical baselines. The original design used a pipeline of specialized sub-agents - a triage agent, an orchestrator, up to fifteen hypothesis investigators and a coding agent - that passed summarized context and voted on verdicts. That architecture was chosen because earlier models struggled to run end-to-end, but it proved costly and brittle: summaries lost crucial context, downstream agents often re-did work, fixes lacked investigative evidence and tended to address symptoms, and debugging required tracing dozens of stage-level logs. Per-agent evaluations masked handoff failures, so individual passes could look good while the overall run failed.

On September 3, a single agent replaced the pipeline and handled triage, investigation and pull-request creation end-to-end. Immediate benefits included one trace per run, faster and higher-quality fixes, and fewer failure modes. Median time from detection to pull request fell from about 2.2 hours to 35 minutes and p90 from days to under two hours; the share of issues ending in PRs rose from 0.6% to over 4% (peaking higher), while cost per PR dropped substantially. Polylane’s conclusions: avoid sub-agent handoffs, evaluate whole runs not individual agents, and keep evidence and action within a single agent.