Skip to content

Agent evaluation: check the work your agents did

See where an agent workflow wasted time or money, which step caused it, and what to change.

Analyse your agent workflows

Tell a wasteful workflow from an efficient one

Two of our own workflows on different inputs: one spent 31.3 credits with an uneven split, the other 16.0 credits with its workers roughly balanced.

Credits per parallel worker: two real workflows of ours, different inputs (agent analysis, 4 Nov 2026)

See the cause, not just the symptom

The slow worker was the symptom; the analysis traced the cause to the step that split the work.
Agent
engineer: why was worker 1 so slow? (example prompt; workflow 1 of ours)
symptom: worker 1 took 436 s; the other workers took 345 s and 225 s
cause, in an earlier step: the step that split the work gave each worker the same number of items, not the same effort
recommended fix: split work by effort, not by count, or process results as they arrive

Your agent gets a recommended fix and can apply it for you

LLM Prover recommends the fix; your agent can propose the edit and apply it with your OK. The exchange shown is an example.
Agent
(example exchange)
finding: uneven work split, caused by the split step -- costs time, about 100.7 s
recommended fix: split work by effort, not by count
agent: I can change the split step to hand out items by expected effort instead of an equal count. Apply this edit?
engineer: yes

Know whether a finding costs time or money

Each finding says which it costs, so you can choose between a faster workflow and a cheaper one.
Agent
engineer: which of these cost time and which cost money? (example prompt; workflow 1 of ours)
time: uneven work split -- the workflow waited about 100.7 s for its slowest worker
money: more turns than needed on worker 1 -- 44 turns against 23 for a matching worker
money: the same documents read by every step -- about 1.07 credits across the five steps
time and money are reported separately and never added together

Analyse a workflow without leaving your agent

Your agent sends the trace (the record of every step) straight from disk, so it never takes up your agent’s working memory.
Agent
analyse my last workflow (example prompt)
> check_trace_eligibility
eligible -- secure upload address returned
> upload the trace file from disk to that address
> analyze_graph_start
analysis running ... done
> analyze_graph
per-step cost and time, each finding, its cause and a recommended fix

Make your agent workflows faster, cheaper and more reliable

See Pro

Questions

Which agents does it work with?
Kiro workflows and Claude Code sessions (Claude Code’s name for one).
What does it cost?
Agent analysis is on Pro and above. If a workflow is larger than your plan allows, the analysis is refused and tells you the numbers.
What happens to my trace?
It is only read as data, never followed as an instruction, and the raw trace is deleted after analysis. Reports are private to you and scrubbed for common key and token formats on a best-effort basis.
Does LLM Prover change my workflow?
No. It gives your agent a recommended fix for each finding; your agent can then propose the edit, and you decide whether to apply it.
What does it check for?
Work done twice, uneven splits, steps that kept going after failures, steps that produced nothing, and where each problem started. Every finding is explained in the Agent Analysis Guide.
How is this different from Agentic Optimization?
Agent evaluation checks work your agents already did. Agentic Optimization lets your agent choose and monitor models.

On Pro: agent evaluation

  • Agent analysis of Kiro workflows and Claude Code sessions
  • Each finding with its cost or time, its cause and a recommended fix
  • Model evaluation for single steps, on the same account
  • Scheduled checks that warn you when a model's quality drops

Analyse workflows from your agent, over MCP (the protocol agents use to call outside tools)

  • Your agent runs the analysis and reads the findings
  • The trace goes straight from disk, never into your agent's working memory

$129/mo $10 for your first week

Then $129/mo. Cancel anytime.

Figures are from agent analyses of two multi-agent research workflows of ours (Kiro, 4 November 2026). The recommended-fix exchange is an example.