Sample private deliverable

AI Agent Cost Leak Report

This sample shows the output format a buyer receives after one repo/workflow audit. Real reports use the buyer's repo, agent logs, and current operating constraints.

Executive readout

High

Exploration waste

Agent repeatedly opens discovery files before landing on the task files.

Medium

Prompt bloat

Project docs mix stable context, old decisions, and task instructions in one long prompt.

Low

Tool loop risk

Test commands are clear, but failure summaries are not compressed before retry.

Top findings

Leak 1 CLAUDE.md describes broad project context but does not give a task-to-file map. First-turn agents spend tool calls rediscovering obvious entry points.
Leak 2 Recent hot files are invisible to the agent. Git activity is a strong clue for maintenance tasks and should be injected in a compact block.
Leak 3 Long test output is pasted raw. A reducer should preserve failing file, test name, stack frame, and assertion while dropping noise.
Leak 4 There is no stop condition for repeated exploration. The agent should switch to asking for missing context after a fixed number of failed searches.

Recommended patch path

Repo map Build a static map of symbols, imports, git-hot files, and package entry points. Inject only top ranked file candidates per task.
Output reducers Compress test, build, stack trace, and JSON output before it reaches the next agent turn.
Runbook Split stable project context from active task instructions, and document the exact verification commands.
View paid sprint