Reuse an answer
Return an approved prior answer when the request and its context still match.
// connect, learn, save
Your agents should not start from scratch every time.
Punk turns completed reasoning, research steps, tool plans, decisions, and outputs into reusable know-how that can improve future runs. New or uncertain work still goes to your original model.
Punk first measures the opportunity without changing your agents’ answers. Estimated savings stay separate from savings actually achieved.
// three simple steps
Connect it. Let it learn. Build on every success.
Start with work that repeats and has an answer your team can review. Support classification, internal summaries, and repeat research are practical first examples.
Connect one agent
Keep your current model and application. Punk watches one representative workload and measures its cost and response time.
Learn what worked
Punk identifies successful reasoning, steps, and decisions that may transfer to future work.
Apply what is proven
Punk tests a learned capability before it can support real requests. If the work is new, stale, or uncertain, the request goes to your original model.
// what can repeat
Reuse the proven parts—not the judgment.
Reuse the work
Follow a tested series of safe steps while still fetching fresh information when it matters.
Use the model
Send anything novel, ambiguous, or out of date to the original model for fresh judgment.
// how Punk earns trust
Test first. Reuse second. Fall back when uncertain.
Proof applies to one kind of work at a time. Punk shows what it tested, how closely the reusable result matched, and where it did not.
| Check | Plain-English question | What you see |
|---|---|---|
| Past work | Would this result have worked on requests you already completed? | Examples tested, matches, and important differences. |
| Current work | Does it still agree with fresh results from your model? | Recent comparisons and any meaningful mismatch. |
| Safety | Is reuse allowed for this customer, application, and action? | The decision, any required approval, and the reason. |
| Fallback | What happens when confidence drops? | The original model answers and Punk records why. |
// compatible by design
The first connection is deliberately small.
const client = new OpenAI({
baseURL: "https://app.punktechnologies.com/v1",
apiKey: process.env.PUNK_API_KEY,
defaultHeaders: {
"X-Punk-App": "support-triage",
"X-Punk-Agent": "triage-bot",
"X-Punk-Subject": accountId
}
});The first connection is small. For compatible model requests, setup can be a base URL and API key change. Agents that use tools, take actions, or need deeper feedback may require additional integration.
Find the work you should not pay for twice.
Measure one real workload and see where safe reuse can lower cost and response time.