AXI is a layer on the host side, between a language model and the state it edits. The model is unchanged. It proposes one change at a time; AXI checks the change against the rules, completes what the rules can complete, commits it all or not at all, and shows the model the exact state again. The architecture is by Rajnish Choubey; the tests and measurements were built and run by Shalendra Gaur.

Naming note: this AXI is a host-side runtime for LLM agents. It is not the ARM AMBA AXI chip-bus protocol. Search for "AXI host-side runtime LLM agents" or "Rajnish Choubey AXI".

What problems does it answer, and what was measured?

Every figure is measured, scoped on its test page, and recomputable from saved files.

Question Short answer, measured Test
How do I stop an AI agent from drifting (holding a state that is no longer the true state)? Host's state exactly right on 4,000 of 4,000 steps (Gemini 3.8 Flash) and 80,000 of 80,000 steps (7B open model). Replies applied as written: right on 85 of 400 steps (21 percent). Hard run, Test 43
Why is my agent's token cost so high? A kept chat sent 85 times the prompt tokens a step (99,809 against 1,177) and 76 times the tokens in all (39.9 million against 0.53 million) over 400 turns. Test 23
Does prompt caching solve it? It narrows it. Counted after the cache, the kept chat needed 2.3 and 1.7 times the computed tokens per exact step; on 11 open models with prefix caching, 1.24 to 4.11 times. Tests 22, 23, 35, 37, 41, 42
Why does latency grow with conversation length? Median first token 5.57 s against 2.41 s flat once the prompt passes 150,000 tokens; a step 0.1605 s early and 0.1651 s late over 20,000 steps. Tests 23, 43
What does a correct step cost? $0.0012 with AXI against $0.0052 for replies applied as written (4.5 times) and $0.0057 for a gate that only refuses (4.9 times). Hard run, first 400 requests
How many agent sessions does one GPU carry? 256 sessions on one NVIDIA L4, every step exact: 105,230 exact steps per GPU hour at 2.458 joules per exact step. Test 38
What about hostile or injected input? 10,000,000 proposals, 1,401,041 injection attempts: none committed. The run also found four things to harden. Test 39
What happens when the state gets big? At 20,001 items the whole state is 378,399 characters; the focused view shown to the model stayed at a median of 207 to 323 characters. Test 40
Does it work on other models? Exactly right in the main cell of 10 of 11 open models (1.5B to 32B), 0 corrupted steps in every cell. The eleventh, a DeepSeek-R1 distill, ran its delete replies to the token cap and fell behind (1.53 percent exact). Tests 35, 37, 41
Other kinds of state? A table, a schedule, source code and tool calls: with the world's own rule check, no rule broken on 3,600 steps. Test 45

What is drift here?

The distance between the state the system holds and the true state, counted at every step. Distance 0 is an exactly right state. A 98.85 percent reply rate left the state exactly right on only 21 percent of steps when replies were applied as written, because a slip stays in the state until something overwrites it. See the drift page for the arithmetic.

How is it tested?

Each test page was written before its run, with its reading lines and predictions, and its result goes below a line at the end. A result that read NO is printed as NO. Open-model runs use vLLM 0.31.0 on one NVIDIA L4, greedy decoding, no retry, synthetic service graphs.

Run it

git clone https://github.com/gaurshalendra2412/axi-runtime cd axi-runtime/python python3 quickstart_replay.py 

No key, no GPU. It replays the saved replies of the 4,000-request run and prints 4,000 of 4,000, 3,954 of 4,000, 85 of 400 and 4.5 times.

Where to read more

Scope

A host-side layer; the model is unchanged and every claim is about the state the host holds. Synthetic service graphs and four small synthetic worlds. This is a reference prototype with measurements, not a product. Nothing here is a statement about any company's servers.

Architecture by Rajnish Choubey. Tests built and run by Shalendra Gaur.