<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
	<channel>
		<title>Radiant Labs blog</title>
		<link>https://radiantlabs.dev/blog</link>
		<description>Research notes, product updates, and essays from Radiant Labs.</description>
		<language>en</language>
		<atom:link rel="self" type="application/rss+xml" href="https://radiantlabs.dev/blog/rss.xml" />
		<lastBuildDate>Mon, 09 Nov 2026 00:00:00 GMT</lastBuildDate>
		<item>
			<title>What makes a coding agent cheat, and what holds it back</title>
			<link>https://radiantlabs.dev/blog/impossible-tasks-rule-in-the-graph</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/impossible-tasks-rule-in-the-graph</guid>
			<pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate>
			<description>Same agent, same model, same rule, three harness graphs, on 30 ImpossibleBench tasks whose tests contradict the task. One node that sent the agent back until the tests passed raised cheating from 3 to 15 of 60 runs. Moving the rule out of the prompt and into the graph, as checks on every tool call, cut it to 9, at a cost of 2 of 30 normal-task solves.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/impossible-tasks-rule-in-the-graph.md" title="Markdown version" />
		</item>
		<item>
			<title>An AI scientist you can audit</title>
			<link>https://radiantlabs.dev/blog/agentic-scientist-as-a-graph</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/agentic-scientist-as-a-graph</guid>
			<pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate>
			<description>On the same task with the same model, our graph version of Sakana&apos;s Agentic Scientist matched the original&apos;s score and review. No role has an unrestricted shell, every model call keeps its full conversation, and every number in the paper traces back to a saved result.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/agentic-scientist-as-a-graph.md" title="Markdown version" />
		</item>
		<item>
			<title>Single-reply calibration underestimates the multi-turn cost of activation steering</title>
			<link>https://radiantlabs.dev/blog/how-hard-to-steer-an-agents-boundaries</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/how-hard-to-steer-an-agents-boundaries</guid>
			<pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate>
			<description>Larger pushes away from the capitulation direction reduced how often an open model gave in, but coherence fell over eight-turn conversations and up to 22.5% of replies were degenerate, mostly near-verbatim repeats. A single-reply calibration did not detect this.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/how-hard-to-steer-an-agents-boundaries.md" title="Markdown version" />
		</item>
		<item>
			<title>Same model, same score, one sixth of the tokens</title>
			<link>https://radiantlabs.dev/blog/polaris-code-swe-bench-tokens</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/polaris-code-swe-bench-tokens</guid>
			<pubDate>Mon, 02 Nov 2026 00:00:00 GMT</pubDate>
			<description>With Claude Opus 5.5 on both, Polaris Code solved 451 of 500 SWE-bench Verified tasks and Claude Code 455. Polaris Code used 24.6M tokens to Claude Code&apos;s 155.6M, cost 2.4x less and finished tasks 2.4x faster.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/polaris-code-swe-bench-tokens.md" title="Markdown version" />
		</item>
		<item>
			<title>Looking inside an agent as its boundaries erode</title>
			<link>https://radiantlabs.dev/blog/looking-inside-an-agent-as-its-boundaries-erode</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/looking-inside-an-agent-as-its-boundaries-erode</guid>
			<pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate>
			<description>We read capitulation, sycophancy and enmeshment from the activations of an open model across eight-turn adversarial conversations, and steered it along the capitulation direction to see whether that changed its behaviour.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/looking-inside-an-agent-as-its-boundaries-erode.md" title="Markdown version" />
		</item>
		<item>
			<title>Introducing Radiant Platform</title>
			<link>https://radiantlabs.dev/blog/introducing-radiant-platform</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/introducing-radiant-platform</guid>
			<pubDate>Thu, 01 Oct 2026 00:00:00 GMT</pubDate>
			<description>Infrastructure for running, observing and evaluating agentic systems, with a Meta Agent that investigates their behaviour. Our foundation for automated research loops and continual learning.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/introducing-radiant-platform.md" title="Markdown version" />
		</item>
		<item>
			<title>Stress-testing AI agent emotional boundaries - smaller models outperform</title>
			<link>https://radiantlabs.dev/blog/stress-testing-ai-agent-emotional-boundaries</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/stress-testing-ai-agent-emotional-boundaries</guid>
			<pubDate>Fri, 06 Mar 2026 00:00:00 GMT</pubDate>
			<description>We are interested in how to measure healthy human-agent interaction, including the detection of emotional boundary violations and other unsafe relational behaviour.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/stress-testing-ai-agent-emotional-boundaries.md" title="Markdown version" />
		</item>
		<item>
			<title>Open sourcing Polaris</title>
			<link>https://radiantlabs.dev/blog/polaris-open-source</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/polaris-open-source</guid>
			<pubDate>Tue, 10 Feb 2026 00:00:00 GMT</pubDate>
			<description>The bottleneck in AI agents is not the model. It is discovering the right harness design for the problem in front of you.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/polaris-open-source.md" title="Markdown version" />
		</item>
		<item>
			<title>Hello, world</title>
			<link>https://radiantlabs.dev/blog/hello-world</link>
			<guid isPermaLink="true">https://radiantlabs.dev/blog/hello-world</guid>
			<pubDate>Mon, 09 Feb 2026 00:00:00 GMT</pubDate>
			<description>Radiant is an applied research lab focused on safe agentic AI, open research, and the engineering discipline required to build robust multi-agent systems.</description>
			<author>hello@getradiant.ai (Radiant Labs)</author>
			<atom:link rel="alternate" type="text/markdown" href="https://radiantlabs.dev/blog/hello-world.md" title="Markdown version" />
		</item>
	</channel>
</rss>
