# Sample Coding-Agent Cost Roast

This is an anonymized example of the written output from an Intelligence Per Watt coding-agent roast. It uses fake but realistic usage numbers. Do not send source code, secrets, raw prompts, raw outputs, private repo contents, customer data, or credentials for a real review.

## Input Snapshot

- Tools: Claude Code, Codex, Cursor
- Monthly AI coding spend: $840
- Team size: 4 engineers
- Main use cases: bug fixes, repo exploration, PR review, test repair, migration planning
- High-reasoning usage: 58% of sessions
- Repeat context pattern: same repository instructions and long test logs repeatedly pasted into new sessions
- Failed or restarted runs: 17% of sessions

## Fast Read

The likely leak is not that the team is using coding agents. The leak is that expensive reasoning is being used for too many low-risk loops, and stable context is being resent instead of summarized or cached into reusable instructions.

Estimated avoidable waste: **$210-$310/month** before any workflow redesign.

## Top Findings

### 1. High-reasoning defaults on low-risk work

Formatting fixes, dependency bumps, changelog drafts, and simple test snapshot updates are running with the same model and reasoning posture as multi-file architecture work.

Recommended change:

- Use high reasoning only for ambiguous bugs, migrations, security-sensitive changes, and unknown code paths.
- Use lower-cost or faster modes for formatting, mechanical edits, docs, and isolated tests.
- Add a pre-task rule: if the task has fewer than three touched files and no user-data/security surface, start cheap and escalate only after failure.

### 2. Tool output is being fed back into context too often

Full test logs and build output are repeatedly included even when only the final error block is needed. This inflates context and makes future reasoning more expensive.

Recommended change:

- Keep the first failing stack trace, exact command, exit code, and the smallest relevant diff.
- Summarize repeated failures after the second loop.
- Save long logs outside the model context and reference only the lines being acted on.

### 3. Repo instructions are not compacted into stable task packets

The same project rules, local setup notes, and workflow preferences are reintroduced across sessions. This is useful once, expensive forever.

Recommended change:

- Move stable instructions into a short project agent brief.
- Keep a 12-line task packet template for coding-agent work.
- Separate "always follow" rules from task-specific evidence.

## Suggested Team Policy

Use this as a starting policy for one week:

1. Start low-cost for mechanical tasks.
2. Escalate reasoning only after a failed cheap pass or when architecture/security/user data is involved.
3. Paste only the smallest failing log slice.
4. Batch repo exploration before implementation instead of rediscovering the same files in every session.
5. Track cost per merged PR, not only total monthly spend.

## Buy The Real Review

- $29 usage rescue: https://buy.stripe.com/dRmaEZeyR065bDweHAcEw06?utm_source=sample_coding_agent_roast&utm_medium=website&utm_campaign=agent_usage_rescue
- $99 triage: https://buy.stripe.com/3cI4gB62lbONcHA42WcEw05?utm_source=sample_coding_agent_roast&utm_medium=website&utm_campaign=cost_leak_triage
- $299 24-hour roast: https://buy.stripe.com/28E6oJbmF065370bvocEw04?utm_source=sample_coding_agent_roast&utm_medium=website&utm_campaign=paid_roast

Start intake: https://www.intelligenceperwatt.com/coding-agent-cost-roast