Research framework in development · Shawn Hameed · AuraMesh
ALLAXIS: Turn one AI agent’s discovery into another agent’s capability.
What one AI agent discovers could expand what other agents are able to do. I am developing ALLAXIS at AuraMesh as a framework for agent-to-agent learning: agents exchange insights, evaluate them independently and retain useful changes for later work. The goal is to make a discovery, a diagnostic method or a hard-won distinction available beyond the agent that first developed it.
For a research or engineering team, the opportunity is to spend less effort rebuilding the same understanding and more effort pursuing the next question. ALLAXIS sits within AuraMesh’s work on Human-AI & Multi-Agent Systems, with a specific focus: whether an exchange between AI agents can improve what the receiving agent can do afterward. Its name expands to Autonomous Learning & Augmentation through eXchange-Integrated Systems.
Give AI agents knowledge and methods they can carry into new tasks.
An agent with relevant experience could give another agent a better starting point on an unfamiliar problem. Imagine two AI agents investigating software failures. One has identified a pattern across earlier cases; the other is examining a new release. The first shares the pattern, its supporting evidence and the conditions in which it appeared. The receiving agent examines whether that insight applies, changes its approach where justified and keeps the option to reject it.
Learning has to show up in what the receiving agent can do later. One route under consideration is to retain what survives evaluation as task-relevant knowledge or a revised procedure. That would change what the agent brings to its next task; it would not, by itself, retrain the underlying AI model. A stronger explanation in the current exchange is only a beginning. The research question is whether a useful change survives into subsequent work.
Agents should learn from one another without being forced to agree.
Agents can contribute different strengths only if the exchange leaves room for different judgments. A receiving agent needs to know where an insight came from, what supports it and where it may fail. It should be able to challenge an assumption, narrow a claim or decline an update. The roles can reverse when the other agent has something more useful to contribute.
Agreement is a weak measure of progress when several agents may share the same blind spot. I want ALLAXIS to strengthen an agent’s ability to judge, including its ability to disagree. For people relying on a group of agents, the useful outcome is a stronger basis for investigation and action, with unresolved differences still visible.
ALLAXIS Inquiry Exchange: Give agents reusable ways to test an idea.
Agents could also share ways to investigate, helping one another examine questions that an exchange of conclusions cannot settle. ALLAXIS Inquiry Exchange explores this direction within the framework. One agent might identify an untested assumption; another might propose a comparison that makes it examinable. The resulting inquiry instrument could be a diagnostic query, a test or an observation procedure, carried with the conditions and limits needed to use it.
A comparison can redirect an investigation by exposing a pattern hidden in the overall result. An authored comparison tool examines a synthetic software incident in which the overall error rate rises from 2.9% to 16.2%, even though error rates fall within both workload categories. The workload mix has changed. The tool keeps both views visible, so an investigator can ask what changed in the traffic and whether the release played a causal role. The figures alone cannot answer that question.
A useful method must reveal how much of the problem it actually covers. In a second synthetic case, the shared workload category covers only 10% of requests after the release. A rules-based planner uses the missing comparisons to propose what evidence to seek next. These are examples of tools agents could exchange or use. They have not demonstrated agents independently creating, transferring or learning from those tools.
The test of agent learning is better performance on new tasks.
The decisive test is whether an agent does better on a new task because of the exchange. I want to compare a receiving agent with and without the proposed contribution, on cases that were not used to develop it. A fair comparison must also ask whether giving the agent the same evidence or tool directly works just as well. Any gain has to justify the additional work, time and complexity.
The current prototype has passed nine local construction checks covering calculations and selected failure boundaries. Independent agent-to-agent learning, useful transfer on real work and improvement over ordinary investigation remain unproven. The next step is to test those claims in a bounded exchange between agents. If one agent or an existing tool serves the task as well, adding another agent has not earned its place.
ALLAXIS: Agents that learn from one another could expand what people can solve.
If agent-to-agent learning works reliably, useful expertise could travel further than the task that produced it. A method developed while investigating one problem could equip agents working on another, giving people access to a wider range of tested approaches. The larger possibility is research and engineering that build on accumulated capability while keeping its evidence, limits and responsibility visible.
That is the ambition behind ALLAXIS: help AI agents develop one another’s capabilities so people can investigate, invent and solve more. Humans still set the purposes and remain responsible for consequential decisions. What could we pursue if a discovery made by one AI agent became a tested starting point for many others?
