Self-Evolving LLM Agents: Compiling SOPs into Tools Reduces Latency by 42%

Researchers published a preprint on arXiv in July 2026 about self-evolving LLM agents: instead of regenerating code for each request, the system pre-compiles standard SOP steps into verified tools. In real deployment on a triage system at a distribution center, this reduced median latency by 42%, and error rates by 53%.

Why ordinary agents waste time

In a standard production agent, the same procedural steps are performed anew with each request: the agent re-invents how to query the database, how to interpret the response, how to make a decision. This wastes time on inference and introduces variability — the same step on two identical requests can produce different code.

The authors propose moving this cycle outside runtime. A special tool-maker runs before deployment: it analyzes the live environment, collects execution traces, studies backend schemas, generates candidate tools and tests them against labeled historical cases. Tools that pass validation are fixed in a versioned registry. The production agent calls them directly and falls back to code generation only when a tool doesn't cover a new situation.

What 1,500 real alerts showed

The approach was tested on a monitoring system at a distribution center: the agent handles production alerts against a 44-node SOP from heterogeneous metric backends.

• Reduction in p50 latency through tool calls: 42% in production
• Reduction in end-to-end error frequency: up to 53% on a sample of 1,500 historical alerts
• Additional reduction in p50 latency with direct-call architecture: another 62% in controlled ablation
• SOP graph size: 44 nodes
• Backends: heterogeneous metric sources

Tools return compact structured verdicts instead of arbitrary text, which simplifies agent architecture: the chain "reasoning → code generation → execution" is replaced by direct tool invocation.

Why versioning matters for operators

Beyond speed, versioned tools solve the transparency problem: every agent decision can be traced to a specific version of a specific tool — critical for production systems where audit is required.

Another benefit is early problem diagnosis. If upstream data changes (data drift), tools begin to fail before users notice. If the SOP contains a specification gap, the tool-maker won't be able to synthesize a tool — and developers learn about it early.

"Self-evolving agents can make industrial LLM systems faster, more

reliable, and easier to operate," the authors conclude.

What this means

The paper proposes a concrete engineering pattern: compile recurring SOP steps into verified tools before deployment. If results reproduce outside logistics, this approach could become a standard for designing production LLM agents — especially where the same analytical process runs thousands of times daily.

Frequently asked questions

What system was the approach tested on?

The agent was deployed in a real triage system at a distribution center: it diagnoses production alerts according to a 44-node SOP, querying heterogeneous metric backends on a sample of 1,500 historical events.

What happens if a ready-made tool fails?

The agent automatically switches to code generation mode (code generation fallback) when an existing tool doesn't cover the incoming request.