Four agent calls became one. Same correct outcomes.
Semio cut coordination input tokens by 75% in a four-scenario KLP test. Both paths got every scenario right. Here is what changed, with the receipts.
Your AI agents should spend their tokens doing the work. Not deciding whether a message is a duplicate, a job is already finished, or a handover refers to yesterday's code.
We put that idea to the test. Semio cut coordination input tokens by 75%. Four model calls became one. Both paths handled all four scenarios correctly.
Same coding artifact. Same evidence. Same coordination model. The difference was what reached the model in the first place.
Three messages did not need another AI decision.
When several agents work together, messages are part of the job. A worker hands over a fix. A reviewer checks it. A coordinator decides what happens next. But not every incoming message needs fresh reasoning.
We compared ordinary agent-managed handovers with the optional Semiotic, or Semio, coordination adapter in Krystal Loop Protocol. We replayed four situations:
- A valid new handover. Both paths used the model to request a bounded review.
- The same handover arriving again. Semio recognised the duplicate without another model call.
- A handover for an old version of the work. Semio rejected the mismatch before it reached the model.
- A receiver reopening after work had already been claimed. Semio's saved attempt record prevented a second launch.
The ordinary agent got those decisions right too. It just needed four model calls to do it. Semio needed one.
Four-scenario coordination comparison
- Agent-managed input tokens
- 5,325
- Semio-backed input tokens
- 1,331
75% fewer coordination input tokens. Four calls reduced to one. 4/4 correct outcomes on both paths.
This was attached to real coding work.
A DeepSeek Flash worker repaired a real notification bug in an isolated copy of KrystalView. An OpenAI critic reviewed the result: different model families, independent roles. All 15 frozen checks and six regression tests passed.
We then used that same completed work for both coordination paths. The valid handover even used a byte-for-byte identical model request. No bloated conversation history. No deliberately confused baseline. No skipped review.
The saving came from removing unnecessary model work, not lowering the bar for the coding.
Structure does the bookkeeping. Models do the thinking.
Semio gives a handover an explicit job, version and message type. The receiving software checks those details against saved records before waking an agent. That turns repeat-work decisions into checks the system can make directly.
For someone building with AI, the benefit is straightforward: your agents can keep working together without spending another model turn on every bit of coordination noise.
This is useful engineering, not a magic symbol. Another well-built runner can implement these checks too; our offline conventional guard matched the filtering decisions. Semio packages them into a reusable, tested adapter that agents and runners can adopt.
The costs, with the cache effect separated
Observed coordination API cost was $0.0169565 versus $0.0021088. Including an equal share of the worker and critic, the totals were $0.0664040 versus $0.0515563.
The Semio call benefited from a warm provider cache. To separate that effect, we also calculated its cost using the baseline's identical input cost. That gives 68.77% lower coordination API cost, or 17.56% lower including the shared coding and review. This is a cache-equalized calculation, not a second live test.
The whole experiment used seven inference calls and $0.1179603 in calculated API usage. The coordinator's subscription cash allocation was unavailable, not zero. The case study separately records preparation and oversight usage.
Where this result applies
This was a controlled four-scenario replay: one new handover and three duplicate, stale or recovery events. The 75% reduction applies to coordination input tokens in that comparison, not every coding token or every production traffic mix. Both paths got the answers right; the improvement was doing less model work to get there.
It is also a separate result from our cheaper-worker coding test. One reduces the price of implementation. This one removes unnecessary coordination calls. Their percentages should not be added together.
Better coordination means fewer wasted model turns. We measured it, kept the checks, and kept the receipts.
Put it to work.
Give your coding agent the Semio adapter guide and ask it to use durable task and version checks before launching another job. Keep KLP's existing scope, review and acceptance rules. No new benchmark is needed to start using the pattern.