AI developers have spent much of the agent boom splitting complicated jobs among teams of specialized agents. Polylane tried that approach in production and eventually found the coordination was getting in the way.
The cloud infrastructure company had built an autofix workflow that could involve up to 18 agents. One handled triage, another coordinated the investigation, as many as 15 tested possible root causes, and a final coding agent wrote the fix.
Then Polylane replaced nearly all of them with one.
In its September 14 account of the overhaul, the company says its single-agent system produced pull requests faster and at a lower model cost. The results offer a useful counterpoint to the idea that more complex agent systems automatically benefit from more delegation.
Polylane’s agents kept losing evidence between handoffs
Polylane built its original multi-agent system in March 2026, when frontier models struggled to take an infrastructure problem from investigation through a code fix. Its own agent harness, the software that manages model execution, was also less mature.
The workflow divided the job into stages. A triage agent assessed an issue, then a coordinator dispatched up to 15 sub-agents to test different root-cause hypotheses. Their findings were summarized for the coordinator, which eventually sent a repair plan to a coding agent.
Each handoff stripped away some of the evidence the next agent needed.
Polylane says the coding agent sometimes received a reasonable repair plan but lacked the original evidence behind it. A summary can preserve the conclusion while dropping observations that would help the next agent challenge that conclusion. The resulting code could address a symptom rather than the actual cause.
Debugging the system created another problem. Engineers sometimes had to inspect dozens of agent traces to understand why one investigation went wrong. Individual agents could perform well on their own evaluations while the complete workflow still failed because information was lost between them.
For developers, Polylane’s experience points to a limitation of evaluating each agent separately. A triage agent can classify an issue correctly, and a coding agent can follow its instructions correctly, yet the final fix can still be wrong if the evidence connecting those steps disappears.
One agent cut the time from detection to pull request
On September 3, Polylane moved triage, investigation, and code generation into one agent. The same model now examines the evidence and carries that context through to the proposed fix.
Here are Polylane’s reported results around the cutover:
Measure | Previous workflow | Single-agent workflow |
|---|---|---|
| Median detection-to-PR time | 2.2 hours | 35 minutes |
| P90 detection-to-PR time | Nine days | Under two hours |
| Detected issues that produced a PR | 0.6% | 4.2%, with further increases reported |
| Average model spend per PR | $111 | About $18 during the first nine days |
Those numbers need some qualification. Polylane says other product changes contributed to the lower costs, so the single-agent redesign cannot claim all of the reduction. A higher pull-request rate also measures proposed fixes, rather than how many fixes were ultimately accepted or resolved the underlying issue.
Even with those caveats, Polylane’s reported results point in the same direction. After replacing the multi-agent workflow with one agent, the company says pull requests were created much faster.
Sub-agents still make sense when work can stay separate
Polylane’s results do not mean sub-agents are useless. Current developer guidance still recommends them for jobs where separate context is useful.
Anthropic recommends Claude Code subagents for tasks that could fill the main conversation with search results or logs. Each sub-agent gets its own context window, custom instructions, and potentially restricted tools.
OpenAI’s Agents SDK also supports specialists that a manager can call as tools, along with handoffs that transfer control to another agent. LangChain’s guidance similarly says one agent can handle many complex tasks, while its subagent pattern remains useful when a supervisor can combine independent results afterward.
The distinction comes down to how dependent each step is on earlier work. Separate research searches or independent code reviews can be split across workers more safely. Polylane’s infrastructure investigations depend on evidence collected early in the process, so compressing that evidence into repeated summaries created more room for mistakes.
Better models can make old agent designs expensive
Polylane also points to another reason to revisit multi-agent systems: the models changed.
Its original design reflected what models could handle in March. By September, Polylane says newer models could complete the entire investigation and repair process themselves.
As models handle more of the workflow themselves, the cost of using multiple agents becomes harder to justify. Every additional agent can mean more model calls, more tokens, and another point where information must be transferred.
A single agent can sometimes cost less per completed task even if each call is more expensive.
The Neuron’s context-engineering explainer examines the same underlying problem: deciding which information a model needs to keep available as work progresses.
Polylane’s experience suggests developers should revisit agent designs as model capabilities improve. If a task depends heavily on evidence collected earlier, adding another specialist may introduce another summary, another model call, and another chance to lose information the final decision still needs.
