We ran an AI agent on a service graph for 400 requests (the first 400 of a 4,000-request run on Gemini 3.8 Flash). 393 of the 400 replies had the right items. But when the replies were applied as written, the state was exactly right on only 85 of 400 steps (21%). With AXI it was exactly right on 400 of 400.

Why 98% right becomes 21% right

A slip that nothing catches stays in the state, and every later step builds on it. Applied as written, the state was off the right one on 315 of the 400 steps. A small error rate per reply becomes a large error rate per job.

What AXI does

AXI is an architecture by Rajnish Choubey. The host holds the exact state of the job. At every step:

1. The host builds a short prompt from the exact state.
2. The model proposes the next change.
3. A deterministic gate checks it (about 2 microseconds, measured in Python).
4. A rule-based repair fixes a slip.
5. The host updates the state.

There is no training and no change to the model. It works from the first step.

What it costs

On the same 400 requests, one step with the right state cost $0.0012 with AXI and $0.0052 when the replies were applied as written (4.5 times). The total spend was about the same ($0.45 to $0.47). The saving is in the cost of a right step, not in a lower bill.

A longer run

In a separate 1,500-step test on Gemini 3.8 Flash, the host-held state was exact on all 1,500 steps, with a prompt of about 2,824 tokens at every step. A chat that kept its whole history had a mean prompt of 476,835 tokens (we stopped it at 457 turns). It was exact for its first 222 turns, then slipped, and applied as written its state was wrong on every turn from turn 287 on.

Don't take our word for it

The test pages were locked with a SHA-256 hash before the runs. A test in the repo recomputes every figure from the saved results. There is a Colab notebook that runs it on your own Gemini key with a spend limit.

What this does not show: it is one run on one model, and the cost figure is from the first 400 requests. A head-to-head comparison against summarisation and other caching set-ups has not been run yet. If you find a number that is wrong, please tell me.

• Repo: https://github.com/gaurshalendra2412/axi-runtime
• Paper: https://doi.org/10.5281/zenodo.23958737

Tests built and run by Shalendra Gaur, published with Rajnish Choubey's permission.

ai #llm #aiagents #tokencost