Use an exact start and end date.
// sample analysis
Know what your agent has learned before you decide what to change.
A Punk analysis shows the current cost, what the agent has learned, which reasoning and processes may be safe to reuse, and what should stay with your model. It ends with a clear recommendation—not a pile of infrastructure metrics.
Illustrative format only
Every bracket below is a placeholder. This page contains no customer result, benchmark, savings claim, accuracy result, or latency result.
// 01 the answer first
[Workload name]: what should we do next?
Recommendation: [keep observing / gather better examples / test one reuse option / activate a small trial / leave this workload with the model].
Separate real traffic from tests.
Name the model and price date.
Name the owner and review date.
// 02 the repeated work
Show the opportunity in language the workload owner recognizes.
| Work pattern | How often it appeared | Possible treatment | What still needs checking |
|---|---|---|---|
| [same request and context] | [count / %] | [reuse saved answer] | [freshness, permissions, quality] |
| [same meaning, different wording] | [count / %] | [reuse when confidence is high] | [edge cases and acceptance threshold] |
| [same process, fresh information] | [count / %] | [repeat the known steps] | [data access and current results] |
| [new or unsuitable work] | [count / %] | [keep original model] | [none; intentionally not optimized] |
Possible savings are not actual savings. At this stage, the original model can keep answering while Punk estimates the opportunity.
// 03 the evidence
Quality comes before the savings estimate.
Past examples
[number tested, comparison method, acceptable results, important failures]
Fresh requests
[number checked silently beside the model, agreement, unresolved differences]
Human review
[reviewer role, sample size, acceptance method, remaining concerns]
// 04 the economics
Keep each number honest.
| Number | Value | What it means |
|---|---|---|
| Current model spend | [$] | Comparable cost for the measured requests. |
| Possible savings | [$] | An estimate while the model still answered; not money already saved. |
| Actual savings | [$] | Cost avoided only after proven reuse answered live requests. |
| Response time change | [ms / %] | A comparable before-and-after measurement. |
| Review or rework | [hours / $] | Human effort reported beside the savings, not hidden. |
// 05 the limits
State what should not change.
Keep with the model
[novel requests, stale information, ambiguous intent, or work without a dependable quality check]
Require stronger control
[sensitive data, consequential actions, permission gaps, or work that needs a person’s approval]
Get a decision, not a dashboard tour.
Bring one workload. The useful outcome is knowing what to reuse, what to keep live, and whether the opportunity is worth pursuing.