Experiment offer Frozen protocol

For Engineering lead at a startup running LLM agents in production (support bots, coding agents, workflow automation) who has at least one recurring agent failure they cannot reproduce

Turn 'it worked when we retried' into a root-cause report

We rebuild one failed agent run from your logs into a deterministic replay with a written root-cause narrative, delivered in 5 business days for $450.

Commercial protocol

The offer, without the pitch deck

Problem
An agent run failed in production, the customer or CEO wants an explanation, and the team cannot reproduce it: prompts, tool outputs, retrieved context, and model responses were not captured in replayable form, so the postmortem stalls at 'it worked when we retried it'
Outcome
Within 5 business days of receiving your logs, you get a replay package for one specific failed run: a deterministic re-execution harness pinned to the captured inputs, an evidence bundle (full prompt/response/tool-call timeline with hashes and timestamps), and a written root-cause narrative you can hand to a customer or auditor
Mechanism
We take whatever partial evidence exists (application logs, provider request logs, trace exports, DB snapshots), reconstruct the run's input state, build a small harness that re-executes the agent loop with recorded tool outputs and pinned model calls stubbed in, and package the timeline plus divergence points where recorded behavior and re-execution differ. Delivered as a repo plus a PDF narrative. Human-built with AI assistance; no access to your production systems required beyond the exported logs
Scope
One failed run per engagement, one agent codebase, up to 50 steps in the agent loop, logs provided by the buyer in any parseable format. Excludes: live production access, fixing the underlying bug (quoted separately), runs where no logs of any kind exist, and ongoing monitoring
Price
$450 fixed per replay package, paid 50% at scoping-call close and 50% on delivery. If we determine after scoping that the evidence is insufficient to reconstruct the run, we say so and refund the deposit in full
First transaction
Buyer books a 30-minute scoping call, shares logs/traces for one failed run under NDA, and pays a fixed fee for one replay package on that single run

Expectation boundary

This page measures interest in the stated offer. Submitting does not create a purchase or guarantee availability.

Declare interest

See whether the protocol fits your operation