Introducing Radiant Platform

Connecting agent execution to investigation and continual improvement.

Radiant Labs

Radiant wave artwork

Today we're introducing Radiant Platform: infrastructure for running, observing and evaluating agentic systems, with a Meta Agent that investigates their behaviour. It brings together the execution record and the tools needed to ask why a run failed, what should change, and how to test that change.

Our starting point is that agents are designs to be discovered. A model's capabilities depend on the system around it: the context it receives, the tools it can call, the information it retains and the conditions under which it revisits a decision. These choices form the agent's harness. Understanding their effects requires experiments and a record of what actually happened.

From execution to investigation

A final answer reveals little about how an agent arrived there. A plausible response can follow an unsuccessful tool call, an incomplete search or a failure to carry information between steps. In a workflow that runs over several days, those failures may be separated from their consequences by many interactions.

Radiant captures execution traces so that model calls, tool use and intermediate steps can be inspected in context. Evaluation and feedback give those traces meaning: whether the task was completed, where behaviour departed from expectations and which failures need investigation. The aim is to connect an assessment of the outcome to the trajectory that produced it.

The Meta Agent works with this recorded evidence. Through tools exposed over the Model Context Protocol (MCP), it can retrieve traces and investigate questions about an agent's behaviour. Its tool calls and results remain available for inspection, so a proposed explanation can be checked against the evidence it used. Project memory provides a place to retain reviewed facts, decisions and recurring patterns across investigations.

This is the foundation we are launching: execution, observability, evaluation and investigation. An explanation of a failure is a hypothesis to test. Establishing that a proposed change improves the system requires another experiment.

Polaris and Radiant

Polaris is our open-source library for composing agentic systems. It makes control flow and the relationships between components explicit, allowing developers to express and revise an agent's design in code. Radiant Platform provides the environment for observing and investigating how those systems behave when they run.

The two address different parts of the same research problem. Polaris gives us a way to express a candidate design; Radiant helps us collect and examine the evidence needed to assess it. We want the connection between design and behaviour to remain visible as systems become more complex.

Closing the experimental loop

Our next step is to connect investigation to controlled harness revision. The Meta Agent would identify a failure pattern, formulate a testable hypothesis and hand an implementation task to a coding agent. That might mean changing a memory policy, revising tool selection or adding a verification step before an action.

The revised harness would then run against a defined evaluation protocol. Development tasks would inform the search, with held-out tasks reserved for validation. Each candidate should retain its parent revision, the proposed change, execution conditions and results, including unsuccessful experiments. Initially, we intend to keep the model and the Meta Agent fixed while varying the harness, so improvements can be attributed to the design under test.

This automated improvement loop is a research direction we are developing. The current platform provides the execution evidence and investigation tools on which it depends.

Continual learning from production experience

Harness revision is one route to improvement. We are also exploring ongoing post-training from agent trajectories: supervised fine-tuning on selected demonstrations, preference optimisation over comparable attempts, and reinforcement learning where task outcomes provide a reliable reward signal.

A production trace needs curation before it becomes training data. We need to know which model and harness generated it, what feedback is trustworthy and which decisions contributed to the outcome. Learning from new tasks also requires measuring whether earlier capabilities are retained. These are research questions for the platform's next phase, alongside the construction of repeatable evaluation and training environments from production workflows.

Our ambition is to make agent development cumulative: each run contributes evidence, each experiment tests a specific claim, and validated improvements inform the next generation of systems. We describe this programme in more detail on the Radiant Platform page.

Work with us

If you're building agents and want to investigate failures, evaluate changes or explore learning from production experience, we'd like to hear from you. Read the documentation or contact us at [email protected].