Your AI agents don't need the whole project. They need the right brief.
Less repeated context. More focus on the job. What we measured with compact packets and the optional Semio adapter in Krystal Loop Protocol.
A capable worker shouldn't need to read eight documents when three contain what the job requires.
That's the idea behind compact context packets in Krystal Loop Protocol: give agents the relevant material, keep additional sources available, and check the result independently.
In plain English: give each AI a clear brief for its part of the job. Don't make it read the entire filing cabinet every time it starts. Keep the important instructions intact, let it ask for more information, and have someone else check its work.
Less reading. The same job to finish.
On one real documentation assignment, compact packets used 20.99% fewer reported input tokens, including review, repair and recheck. Both approaches passed the final checks.
Reported input tokens, including repair and recheck
- Full context
- 104,544
- Compact packets
- 82,601
One paired assignment, measured by KU on 19 October 2026.
The full brief included eight reference documents. The compact brief selected three relevant documents from the same set. Both approaches could request more information; neither needed extra reads. Each went through one repair and a further review.
That last detail matters. The first reviewers missed the same ambiguous count. The coordinator caught it, and the repair is included in the totals. This was a checked result, not a claim that any agent or reviewer is infallible.
Stop asking "is it ready yet?"
Our optional Semio adapter tackles another source of waste: repeated checking while waiting for work. In a separate controlled comparison, transport reads fell from 21 to 3.
Think of it as receiving a notification when something is ready instead of repeatedly checking the inbox. That is fewer transport reads, not a measure of token, money or time savings. Neither waiting method called a model, and the prompt size stayed the same.
Clearer messages, not magic symbols
Messages identify the job, the revision and the kind of update explicitly. The receiving software checks those details. That reduces ambiguity about which piece of work a message belongs to.
Symbols do not force a model to reason correctly. Structured messages can still contain wrong information. The benefit comes from shared structure and validation, alongside the existing checks on the work itself.
What we can honestly claim
These results do not establish universal cost savings, faster delivery or better accuracy. They show specific reductions we can explain and inspect.
These are KU's own observations, not an independent audit. Worker and reviewer models count tokens differently, and coordinator usage and monetary costs were not fully available. A reduction in reported input is not automatically the same reduction on a bill. The two comparisons measure different things; their percentages cannot be added together.
The ambition is straightforward: spend less effort repeating the brief, and more effort getting the work right.
Try it with your coding agent
Give your agent the repository and ask how it would apply KLP to one bounded task. Compact packets and Semio are optional; you don't need either to use the core protocol.