XAGI Labs is a frontier AI research lab building the infrastructure required to give autonomous AI real power safely.

AI systems can increasingly browse, write code, operate computers, move money, modify production systems, and execute multi-step workflows.

The problem is no longer just what AI can do.

It is:

• What is it allowed to do?
• What actually happened?
• Can we prove it?
• How much autonomy should this system be trusted with?

XAGI sits between AI agents and the systems they act on, enforcing bounded authority, independently verifying outcomes, and producing evidence that can be used to assess how much autonomy an AI system should receive.

The model does the thinking. XAGI does the allowing, the proving, and the undoing.

Our work follows a simple loop:

MELRA is XAGI's open-source operating layer for governed AI execution.

It sits between agents and real systems and provides a common control path across:

• Files
• Terminal
• Browser
• Computer use
• Memory
• External tools

Actions pass through policy evaluation, scoped permissions, approvals, verification, receipts, and execution certificates.

MELRA is:

• Model-agnostic
• Agent-framework agnostic
• Designed to run inside the customer's environment
• Open source under Apache 2.0

→ Explore MELRA

SOVA is our work on verifiable evidence for autonomous systems.

A .sova capsule describes a behaviour, its operating context, required evidence, trials, and limitations.

A .sova-trace records observable events and their causal relationships so claims about an AI system can be evaluated against what actually happened rather than what the agent or tool provider says happened.

The goal is simple:

Don't trust "success." Prove the effect.

TRUSCOR is XAGI's rating and assurance layer.

Using evidence collected from deployed systems and adversarial evaluations, TRUSCOR is being developed to answer a harder question:

How much autonomy can this AI system actually be trusted with?

The long-term objective is to turn AI behaviour from something organisations merely observe into exposure that can be measured, compared, rated, and eventually priced.

Enterprise AI increasingly fails at the same boundary.

The agent works.

Then someone asks:

What is the worst this system can do, how do we limit it, how would we prove what happened, and who carries the risk if it goes wrong?

Today, those questions rarely have satisfactory answers.

We believe autonomous AI needs the same infrastructure that other high-consequence technologies eventually developed:

a limit, a record, and a rating.

That is the layer XAGI is building.

Independence matters.

XAGI does not build:

• AI agents for customers
• General-purpose agent applications
• Agent frameworks
• Business workflows that make decisions for users
• Remediation products based on our own findings
• General-purpose AI observability tools

We also do not rate systems built by XAGI itself.

A company selling the thing being evaluated has an obvious interest in the answer.

We intend to build the independent layer underneath the ecosystem instead.

Our work currently spans:

• Autonomous AI accountability
• Policy-governed agent execution
• AI permission and authority systems
• Verifiable execution and evidence
• Adversarial AI evaluation
• AI autonomy and trust ratings
• Agent risk measurement
• Computer-use safety
• Evidence provenance and causal tracing
• AI assurance infrastructure
• Long-running agent reliability
• Machine-generated evidence standards

Our engineering and research are developed with a simple rule:

If we haven't measured it, we don't claim it.

Negative results, incomplete evidence, and failed experiments are part of the record too.

We believe critical infrastructure for AI accountability should be inspectable.

MELRA is open source, and parts of our evidence architecture and research are being developed publicly so developers, researchers, security teams, and enterprises can inspect the assumptions rather than simply trust our marketing.

Because apparently building systems that ask people to trust the trust system would be a rather spectacular design failure.

Today:

Bound AI → Record what happened → Rate autonomy

Over time:

Measure exposure → Build loss history → Make AI risk priceable and transferable

The ambition is not to build another agent.

It is to build the infrastructure that allows an entire ecosystem of agents to operate with real authority without asking organisations to operate on blind trust.

We're interested in working with:

• AI infrastructure engineers
• Security researchers
• Agent developers
• Enterprise AI teams
• Red-teamers
• Evidence and provenance researchers
• Risk and insurance researchers
• Open-source contributors

Contributions to our public repositories are welcome.