Polylane says one agent improved quality while cutting latency and cost

• Polylane says replacing its orchestrated sub-agent workflow with one end-to-end agent improved PR quality while reducing latency and cost, according to boristane’s announcement.
• The prior Autofix design could use 15 sub-agents plus an orchestrator, with a confirmed issue involving as many as 18 agents, according to Polylane’s case study.
• A time-to-PR chart shared by boristane marks the September 3 switch to one agent and ends at a 30-minute daily median after earlier results ranged as high as two weeks.
• The reported gains are production observations, not an architecture-controlled benchmark. In boristane’s benchmark reply, the author said separating the effects would require another benchmark, which the team probably would not run.
• The company still distinguishes independent and background workers from the handoff-heavy pattern it calls sub-agents, as boristane’s terminology makes explicit.

Polylane’s current product page describes an agent that reads code, watches infrastructure and opens fixes as pull requests. Its case study puts a previous incident-fixing path at 15 sub-agents and an orchestrator; teortaxesTex called the disclosure “impressive and commendably open,” while boristane said the prompts were not secret but were too large to post.

The old Autofix pipeline split an incident into triage, coordination, multiple hypothesis investigations and a coding agent that submitted the PR. The case study says its orchestrator could fan work out to 15 sub-agents, then receive summaries rather than the full evidence trail.

Polylane attributes three failures to that design:

• Handoff summaries lost context, so agents repeated work and the eventual PR could treat a symptom rather than the root cause.
• Each specialist saw only part of the investigation, while the coding agent inherited a compressed version of it.
• Debugging meant following many traces and evaluating stages individually, without a reliable end-to-end quality signal.

Questioning did not restore much of that missing context. In a boristane reply, the author said the main agent could ask sub-agents more questions but “rarely did.”

Polylane says it removed the triage agent, coordinator, hypothesis workers and coding handoff on September 3. One agent then handled triage, investigation and PR submission in a single trace, according to the case study.

The chart tracks hours from issue detection to opening a PR. Its daily median and median-to-p90 band fluctuate sharply before the cutover, then settle near the bottom of the log-scaled axis, with the final value labeled 30 minutes.

A separate post adds an important bit of context: boristane said a teammate who had joined four weeks earlier had rewritten half of Polylane and published two engineering posts. The time series therefore captures a production-system change made amid broader engineering work, not a clean A/B test of delegation.

The public account supplies a timing series, but it does not allocate the observed quality, cost and latency changes among model behavior, the rewritten workflow and other system changes. another boristane post summarized the thesis as “quality inference is the whole game,” without publishing a corresponding quality metric in the post.

The prompt set remains unavailable. Boristane said it was “nothing secret” but too large to post, limiting the material available to reproduce the single-agent harness. In a separate boristane post, the author also made an adoption claim without naming the teams or giving a customer metric.

Polylane is not arguing that every other agent disappears. boristane’s terminology defines “background agents” as workers fed by a user-facing main agent that can autonomously update the user or each other.

In that vocabulary, a sub-agent is the narrower Claude Code-style pattern: parallel workers send a short summary back to the lead. The distinction tracks zeeg’s clarification, which says the team still has other independent agents but avoids orchestrator fan-out by doing the main work inline.

A reply from thdxr makes the counter-case concrete: a Slack-like, multi-user interface can benefit from background work that handles concurrent requests. Polylane’s result is a case against its previous handoff chain, not evidence that concurrency is universally harmful.