Open protocol from Krystal Unity

Krystal Loop Protocol

Hard-working agents. Big-brain control. No chaos.

Let fast, affordable AI agents do the heavy lifting while a stronger lead keeps the job focused, checks what comes back, and proves the product still does its main job when the work is finished.

Why it exists

More work done without losing the plot.

Agent loops can move incredibly fast. They can also burn through your best model's attention, wander away from the original job, or confidently tell you that broken work is finished while quietly removing features that worked before.

KLP freezes the product promise before work begins. The team defines the job, lets a worker build it, checks the result, gets a fresh second opinion, and repairs only what the evidence supports. The loop cannot finish until the product's important capabilities are proven again.

If the evidence shows the product will fail at a job it promises to handle, the team cannot call it finished just because the failure has not happened in production yet. Hypothetical edge cases do not automatically block completion or send agents into endless repair loops.

The loop

Six clear stages. One protected outcome.

Each stage has a purpose, no agent marks its own homework, and green technical checks cannot hide a broken product.

01

Promise

Record what the product must keep doing and the proof that it works now.

02

Scope

Define the outcome, boundaries, budget, checks, and stop conditions.

03

Build + check

Give a capable worker one bounded job, then run direct checks on its result.

04

Critique + repair

A fresh agent reviews the exact result; verified problems get bounded repairs.

05

Integrate

Treat the combined system as a new result and check it again.

06

Prove

Re-run the product proof. A lost or unproven core capability stops completion.

What changes

Vibe coding without the hangover.

Spend the big brain wisely

Keep stronger models focused on planning and review. Give suitable bounded jobs to a lower-cost capable worker. Model choice can reduce cost, but review and repair effort still count.

Keep every agent on the same job

Give each agent the context its assignment needs, not the entire project history. Keep the full job requirements and evidence. Explicit task IDs and revisions make handover mismatches easier to catch.

Stop loops becoming endless

Set one shared effort budget and a repair limit. Include extra reads, reviews and retries, then stop or reassess when the agreed limits are reached.

Measured on 10 November 2026

As many tasks solved. A sixteenth of the cost off-peak.

Two public benchmarks on real open-source code, graded by tests the models never see. Every figure recomputes from receipts in the repository.

59 of 66 tasks at $0.046 each

A low-cost model built each feature and a frontier model reviewed the exact result. The heavy model working alone solved 56 at $0.721 each, and the frontier model alone solved 56 at $0.123. The critic confirmed 77 findings by running a failing test, and 70 were fixed before handover.

28 of 30 conflicting pairs at $0.103

Two KLP workers, joined by the Semiotic Protocol (Semio), delivered 28 of 30 pairs of conflicting features with both test sets passing. The frontier model building both alone delivered 22 at $0.196 a pair. Against agents coordinating by chat, Semio reached the same 28 with 333 model calls in place of 581.

75% fewer coordination input tokens

In a four-scenario comparison, Semio turned four coordination model calls into one and cut input tokens from 5,325 to 1,331, with every scenario handled correctly on both paths.

21% fewer input tokens with compact briefs

On one real documentation assignment, compact packets used 82,601 worker and critic input tokens where the full reference bundle used 104,544, repairs included. Waiting for a handover by event took 3 update reads where polling took 21.

Clear messages, not magic symbols.

The optional Semio adapter labels what a message means and which task and revision it belongs to. Receivers check those fields. That reduces ambiguity about the job being discussed, but it does not make the message true or give an agent permission to act. The wake bridge starts a preconfigured job; it does not use message text as launch instructions. This is a bounded protection, not prompt-injection immunity.

Current release

One command, with a receipt.

Runner 0.5.0 is public and provider-neutral. klp run builds the change in a sandbox, runs your checks, has an independent critic review the exact result and hands you a patch and a receipt. For changes to saved data, security or money, KLP Core protects a frozen Mission Contract and checks the product promise before and after the work.

Release Runner 0.5.0, KLP Core v0.3.3 draft
Models Provider-neutral
Proof Two public benchmarks, recomputable from receipts
License Free for individuals and small businesses

Give it to your AI coding agent.

Ask what it would change about the way your agents plan, build, check, and repair work today.