Agent tracing: find where workflows waste time and money

This page is for anyone who runs AI agents that work through several steps. It gives six waste patterns you can find in a workflow’s trace, the fix for each, and how an agent analysis finds them automatically.

Terms used below:

AI agent: software that uses an AI model to work through steps on its own.
Trace: the record of a workflow, with every step, what it cost and how long it took.
Agent analysis: LLM Prover’s step-by-step check of a workflow’s trace.
Turn: one time the model thought and decided.
Tool call: an action the agent took, such as a read, fetch, search or edit.
Worker: a step that does the main work with tools; in the workflow below, three workers ran side by side.
Credits: the unit our workflows are billed in.

The six patterns and their fixes

Pattern	How it shows in a trace	Fix	Costs
Uneven work split	One worker takes much longer than the others	Split work by effort, not by count (where possible), or process results as they arrive	Time
Same documents read by every step	The same files appear in every step’s input	Load shared documents once and share them	Money
More turns than needed	One step’s turn count is much higher than an identical step’s	Tighten the instruction or add a stop condition	Money
Kept going after repeated failures	Failed tool calls followed by more of the same	Flag it back to the assigning agent, or submit what it found so far	Money and time
Step produced nothing	A step spent credits but returned no usable result	Check the step’s input and stop conditions	Money
Same call repeated	Identical tool calls with identical inputs	Reuse the earlier result	Money and time

The figures below come from an agent analysis of one of our own multi-agent research workflows, analysed on 4 November 2026; the workflow cost 31.3 credits. That workflow shows the first three patterns; the other three are described as patterns only.

Workflow 1, per step (agent analysis output, analysed 4 November 2026)

Build and split the list: 1.07 credits, 43 s, 7 turns, 8 tool calls
Worker 1: 12.05 credits, 436 s, 44 turns, 111 tool calls
Worker 2: 11.26 credits, 345 s, 43 turns, 94 tool calls
Worker 3: 4.97 credits, 225 s, 23 turns, 63 tool calls
Check and record: 1.96 credits, 102 s, 11 turns, 22 tool calls

Split work by effort, not by count (where possible)

The first step gave each worker the same number of items, regardless of how much work each needed. Worker 1 took 436 seconds and worker 3 took 225, and the agent analysis put the cost of the uneven split at about 100.7 seconds of waiting compared with an even split. It traced the cause to the step that made the split.

Spot it: compare the time each parallel worker took.
Fix: divide the work by expected effort where you can (difficulty is not always known up front), let finished workers pick up remaining items, or have the next step process results as they arrive.
Costs: time. The agent analysis reported it as a time cost, not a money cost.

Each of the five steps read the same two documents on its own: about 21 KB per step and about 105 KB across the workflow. The agent analysis put this at about 1.07 credits.

Spot it: look for the same files in the input of every step.
Fix: read shared documents once and pass the content, or a summary, to the steps that need it.
Costs: money.

Tighten the instruction or add a stop condition

Workers 1 and 2 took 44 and 43 turns, against 23 for worker 3. The agent analysis judged the extra turns wasteful on worker 1, but not on worker 2, whose share may have held harder items.

Spot it: compare turn counts across steps given the same job, and check whether the busier step simply had harder input.
Fix: make the instruction more specific, with a clear stop condition, or spread hard items across steps.
Costs: money. Every turn is paid for.

Insight: In one workflow of 31.3 credits, the agent analysis found an uneven split, repeated reading and extra turns on one worker. Each has a specific fix.

Make your agent workflows faster, cheaper and more reliable

An agent analysis checks every step of a workflow and shows where it wasted money, time or effort, and what caused it. On Pro and above.

See Pro

Flag repeated failures instead of carrying on

A step can keep trying after its tool calls keep failing. For example (illustrative), a step given links that no longer work may try every one of them.

Spot it: look for several failed tool calls in a row, followed by more attempts of the same kind.
Fix: have the step flag it back to the agent that assigned the work for review, or do a quick review of what it has and submit that if it cannot continue.
Costs: money and time.

Check a step that produced nothing

A step can spend credits and return no usable result. That is only waste if the step failed: if it searched properly and there was nothing to find, that is an honest result.

Spot it: look for a step with cost but no usable output.
Fix: check the step’s input and stop conditions; if it failed, fix the input or add an early exit.
Costs: money.

Reuse the result of a repeated call

The same tool call with the same inputs can appear more than once in a workflow. Identical calls return the same result.

Spot it: look for identical calls with identical inputs.
Fix: reuse the earlier result, or tell the step to keep track of what it already fetched.
Costs: money and time.

How an agent analysis finds these automatically

An agent analysis treats a workflow as a map of its steps and which step feeds which. It works in two passes:

An automatic pass walks that map the same way every time and flags likely problems, such as the patterns above.
Some flagged problems then get a reasoned review: what was the step for, did it do that, and does the cause sit in an earlier step?

Each finding that affects cost or time says which, and the two are never added together.

You run it from your agent over MCP (the protocol agents use to call outside tools), on Kiro and Claude Code workflows. Your agent uploads the trace file straight from disk to a secure upload address, so the trace never takes up the agent’s working memory. What each label in the report means: Reading an agent analysis.

Insight: A trace shows what happened. An agent analysis turns it into a list of fixes, each with the step, what it costs and why.

Questions

What is the difference between tracing and an agent analysis? Tracing records what happened in a workflow. An agent analysis measures the workflow one level up from a single model call: it checks each step of the whole workflow, and reports where work was wasted and which step caused it.

Does uploading a trace load it into my agent’s working memory? No. Your agent sends the file straight from disk to a secure upload address.

Analyse your agent workflows from your agent over MCP and get the list of fixes for each workflow, on Pro and above: see how agent evaluation works.