Repeated work
Requests or processes that are similar enough to work your agent has completed before. Repetition is an opportunity, not automatic permission to reuse.
// plain-English guide
You should not need a glossary to understand Punk.
Punk turns completed reasoning, research steps, tool plans, decisions, and outputs into reusable know-how, then proves when it can safely improve future work. This page explains the few supporting ideas without turning them into a new category.
// the essentials
Four ideas explain the product.
Proven work
A saved result or known process that has passed the quality, scope, freshness, and safety checks required for a defined workload.
Return to the model
When reuse is not allowed, not relevant, or does not work, Punk sends the request to your configured model.
Actual savings
Model cost avoided after proven work really answered a live request. This is reported separately from possible future savings.
// how reuse earns trust
Test first. Use second. Keep measuring.
Test against past work
Run the proposed reuse on recorded examples and compare it with answers the team already accepts. This shows how it handles known cases.
Test beside live work
Try the proposed reuse silently on fresh requests while the original model still supplies the answer users receive.
Require evidence before use
Define the quality and safety bar a reuse option must meet before it may answer live requests.
Start small
Limit a newly approved option to a small, known share of suitable requests while Punk measures real behavior.
// kinds of reuse
Different work can repeat in different ways.
The same request
Reuse a saved result when the relevant request, customer, permissions, and freshness rules match exactly.
The same meaning
Consider a saved result when two requests mean the same thing, even if their wording differs. This needs stricter confidence and review.
The same information lookup
Reuse an earlier read when its inputs, permission scope, and freshness still match. Actions that change the outside world are treated differently.
The same process
Repeat a tested sequence of steps while still fetching any information that must stay current.
Possible savings
An estimate made while the original model still answers. It shows opportunity, not money already saved.
Realized savings
Measured model cost avoided by proven work that actually answered live requests.
// terms you may see in technical docs
Older and engineering language, translated.
| Technical term | What it means in plain English |
|---|---|
| Adaptive runtime | An earlier category label for Punk’s learning optimization system. We no longer lead with it. |
| AI gateway | The compatible request interface Punk uses to receive model traffic; compatibility, not Punk’s main product promise. |
| Agent observability | Tools for inspecting what an agent did. Punk provides focused evidence for reuse but does not aim to replace every monitoring product. |
| Agent governance | Rules that control who and what an agent may access or change. Punk honors those rules during reuse. |
| Artifact | An internal name for a tested, reusable process. |
| Route explanation | The stored reason Punk reused prior work or sent the request to the model. |
| Trace | A record of what happened during one agent request. |
| Shadow or replay | Engineering names for testing on fresh or past work without letting the test change the outside world. |
The product is simpler than the terminology.
Punk learns from completed work, proves what can transfer, and leaves everything else with your model.