Independent project · Concept stage

A better
next attempt.

Your agent executes. XELON asks what should happen next.

A proposed local-first improvement protocol for existing AI agents. Evaluate the result. Learn from the evidence. Change the strategy.

Explore the loop Read the founding thesis

An idea to test. Not a proven performance claim.

X → E → L → O → N

Progress needs evidence.

A loop with a reason to continue.

XExecute

EEvaluate

LLearn

OOptimize

NNext

EXECUTE / EVALUATE / ADAPT

Bring your own agentBring your own modelBring your own compute

The starting point

Execution is an attempt. Success needs a definition.

A plausible answer can still miss a constraint. Another attempt can repeat the same mistake. XELON proposes a bridge between doing the work and knowing what to change.

A clear goal

Define the outcome and acceptance criteria before the agent starts.

An honest evaluation

Check the artifacts against those criteria. An agent saying “done” is not evidence.

A justified next step

Change the strategy when the evidence supports it. Stop when it does not.

Inside the protocol

Five stages. One deliberate cycle.

Select a stage to inspect an illustrative authentication-system task. This is an interactive explanation, not a running XELON framework.

Execute

Turn a goal into an attempt.

An existing agent works on the task using the tools and permissions the user provides.

Attempt artifact

The agent adds sign-in, session handling, and authentication tests. The output now needs independent checks.

The complete stage explanation is available below; JavaScript enables the interactive controls.

Evaluate — Measure against the goal.

Evaluate the result against declared criteria. Preserve failures and uncertainty instead of trusting a completion message.

In this example, tests expose a session-expiry failure. A specified security check flags missing cookie settings. Compatibility and build checks pass.

Learn — Extract something usable.

History says what happened. A structured learning connects an observation to evidence and its limits.

Session expiry does not match the test requirement. Cookie settings need review. These observations concern this attempt; they are not general performance findings.

Optimize — Change the next strategy.

Use the learning to choose a specific change. A stored transcript alone does not improve the next attempt.

Focus the next attempt on expiry logic and cookie configuration, then rerun the complete check set to catch regressions.

Next — Continue for a reason. Or stop.

Check the iteration, time, and compute limits before another attempt. Stop on success, exhausted limits, or a need for human input.

If budget remains, authorize the revised attempt. Otherwise stop and report the unmet criteria. No improved outcome is assumed.

The difference that matters

Retrying is not the same as improving.

Evaluation is the center of the idea. Without a trustworthy signal, a loop can get longer without getting better.

History

“The session test failed.”

A record of an event.

Learning

“Expiry behavior contradicts the requirement.”

An observation linked to evidence.

Optimization

“Revise expiry logic; rerun all checks.”

A concrete change to the next strategy.

Your stack. Your boundaries.

Bring the intelligence. Keep the control.

XELON is intended to work around the agent you choose. It is not a model, a hosted agent, or a replacement for your existing system.

Local-first

Configuration, artifacts, and iteration records should stay local by default. Remote model calls may still send data externally.

Agent & model agnostic

Adapters are a proposed integration approach. No working integrations or supported models are claimed yet.

Bounded by you

Set goals, permissions, iteration limits, and compute budgets. More attempts should always have a reason.

The research question

Does XELON make agents measurably better?

We do not know yet. The proposed benchmark compares the same agent with and without XELON under comparable conditions and equal budgets.

Agent alone

Its normal workflow, including its existing retry behavior.

Agent + XELON

The proposed evaluation, learning, and optimization cycle.

Task success
Output quality
Elapsed time
Token / compute cost
Failure analysis

No benchmark results yet. Negative and inconclusive outcomes count as evidence, too.

Read the validation plan

An open direction

Build the evidence. Then build on it.

A free, open-source core is the intended direction. No license has been granted and no framework package is available. Coding agents are a first potential testing ground, not the permanent scope.

npx xelon initPossible future CLI · Not implemented

XELON / FOUNDING RECORD v0.1

Follow an idea from its first principles.

Start with the published founding record. It defines the hypothesis, proposed architecture, and what still needs to be tested.

Read Founding Record v0.1No signup form or mailing list is connected. The repository is private; the links below open the founding documents on this site.