# Oxagen Research: the science of code graphs, agents, and self-improving systems Source: https://oxagen.sh/blog Research # What the research says about agents, and how to govern them Research notes on code graphs, AI agents, coding agents, self-improving models, self-evolving agents, and autonomous agents. Each note covers what the papers show, where the methods fail, and what it takes to govern the agents built on them. Pillars ## Research topics Every post belongs to at least one pillar. Open a pillar to see all of its posts. [ ### Code graphs Why agents work better when facts are linked. 3 posts ](https://oxagen.sh/blog/pillars/ontologies)[ ### AI agents Models that take actions, and what makes them reliable. 11 posts ](https://oxagen.sh/blog/pillars/ai-agents)[ ### Coding agents Agents that read, write, and fix code. 7 posts ](https://oxagen.sh/blog/pillars/coding-agents)[ ### Self-improving models Models that train on their own output. 2 posts ](https://oxagen.sh/blog/pillars/self-improving-models)[ ### Self-evolving agents Agents that change their own skills and code. 2 posts ](https://oxagen.sh/blog/pillars/self-evolving-agents)[ ### Autonomous agents Agents that work for days without a person watching. 8 posts ](https://oxagen.sh/blog/pillars/autonomous-agents)[ ### Product Releases What shipped in each Oxagen release and why. 8 posts ](https://oxagen.sh/blog/pillars/product-releases) Latest ## All posts [](https://oxagen.sh/blog/a-prepaid-wallet-for-agent-spend) * [Product Releases](https://oxagen.sh/blog/pillars/product-releases) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [A prepaid wallet for agent spend](https://oxagen.sh/blog/a-prepaid-wallet-for-agent-spend) Oxagen 5.5 prices each model call by size before it runs, pays from a monthly allowance and then a prepaid wallet, and holds new runs when the wallet is dry while started runs keep their rules. November 5, 2026 · 5 min read [](https://oxagen.sh/blog/cloud-sandboxes-for-coding-agents) * [Product Releases](https://oxagen.sh/blog/pillars/product-releases) * [Coding agents](https://oxagen.sh/blog/pillars/coding-agents) ### [Cloud sandboxes for coding agents](https://oxagen.sh/blog/cloud-sandboxes-for-coding-agents) A coding agent on a laptop stops when the lid closes or the plan's usage limit hits. Oxagen runs it in a sandbox on your Modal account, with every linked repository and a terminal that pauses it. November 5, 2026 · 5 min read [](https://oxagen.sh/blog/decisions-in-one-place-the-console-and-the-work-board) * [Product Releases](https://oxagen.sh/blog/pillars/product-releases) * [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents) ### [Decisions in one place: the Console and the Work board](https://oxagen.sh/blog/decisions-in-one-place-the-console-and-the-work-board) Agents raise questions, CI fails at night, and results pile up for review. The Console asks each decision as a choice, and the Work board turns a drag into the action it means. November 5, 2026 · 6 min read [](https://oxagen.sh/blog/evidence-an-auditor-can-check) * [Product Releases](https://oxagen.sh/blog/pillars/product-releases) * [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents) ### [Evidence an auditor can check](https://oxagen.sh/blog/evidence-an-auditor-can-check) An auditor should not have to trust the vendor that kept an agent's record. Oxagen 5.5.0 checks evidence offline, maps AIUC-1 control by control, and scores operators from the record. November 5, 2026 · 5 min read [](https://oxagen.sh/blog/model-credentials-that-follow-the-person) * [Product Releases](https://oxagen.sh/blog/pillars/product-releases) * [Coding agents](https://oxagen.sh/blog/pillars/coding-agents) ### [Model credentials that follow the person](https://oxagen.sh/blog/model-credentials-that-follow-the-person) A key saved once serves every organization you work in. A run moves to your next saved login when one runs out, and a refused login asks for a new sign-in on your own machine. November 5, 2026 · 6 min read [](https://oxagen.sh/blog/oxagen-5-5-0) * [Product Releases](https://oxagen.sh/blog/pillars/product-releases) ### [Oxagen 5.5.0, the last beta](https://oxagen.sh/blog/oxagen-5-5-0) Oxagen 5.5.0 is the last beta, a pre-production release. It runs coding agents in cloud sandboxes, prices model calls before they run, and puts every decision in one place. November 5, 2026 · 6 min read [](https://oxagen.sh/blog/reading-a-long-agent-run) * [Product Releases](https://oxagen.sh/blog/pillars/product-releases) * [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents) ### [Reading a long agent run](https://oxagen.sh/blog/reading-a-long-agent-run) A long run hides its failures in hundreds of steps. The Run page now lists the run's loops, refusals, setup failures, and failing tools, and tells a wrapped agent when it loops. November 5, 2026 · 6 min read [](https://oxagen.sh/blog/steering-drafted-from-corrections) * [Product Releases](https://oxagen.sh/blog/pillars/product-releases) * [Coding agents](https://oxagen.sh/blog/pillars/coding-agents) ### [Steering drafted from corrections](https://oxagen.sh/blog/steering-drafted-from-corrections) Oxagen now drafts steering from what people correct mid-run, reads refusals and failing tools across runs each night, and serves steering only to the agents it wraps. November 5, 2026 · 6 min read [](https://oxagen.sh/blog/agents-are-waiting-on-a-process-built-for-people) * [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents) * [Coding agents](https://oxagen.sh/blog/pillars/coding-agents) ### [Agents are waiting on a process built for people](https://oxagen.sh/blog/agents-are-waiting-on-a-process-built-for-people) An agent can write a change in minutes. Then the change waits for a person to read it. This post measures that wait, what eight companies changed about it, and how Oxagen ships at every hour. September 30, 2026 · 11 min read [](https://oxagen.sh/blog/steering-a-run-you-are-not-watching) * [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [Steering a run you are not watching](https://oxagen.sh/blog/steering-a-run-you-are-not-watching) When a long run goes wrong, most teams can stop it or type at it. Both work badly. A steer is a third option. It is a message with a delivery mode, a status, and a record. September 16, 2026 · 6 min read [](https://oxagen.sh/blog/the-agent-time-horizon-is-doubling) * [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [The agent time horizon is doubling](https://oxagen.sh/blog/the-agent-time-horizon-is-doubling) METR measures how long a task an agent can finish on its own. That length has doubled about every seven months since 2019. This post covers what week-long runs mean for supervision. September 16, 2026 · 6 min read [](https://oxagen.sh/blog/the-problem-is-not-slop-it-is-your-process) * [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [The problem is not slop, it is your process](https://oxagen.sh/blog/the-problem-is-not-slop-it-is-your-process) The worry about AI slop is about output quality. The measurements point to a different cause. Teams give an agent a workflow built for people and expect it to work. September 16, 2026 · 6 min read [](https://oxagen.sh/blog/what-an-agent-should-be-told-before-it-starts) * [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [What an agent should be told before it starts](https://oxagen.sh/blog/what-an-agent-should-be-told-before-it-starts) A prompt file has no owner, no date, no scope, and no record that the agent read it. So it is a poor place for a standing rule. Steering records give each rule those things. September 16, 2026 · 7 min read [](https://oxagen.sh/blog/the-limits-of-self-correction-and-model-collapse) * [Self-improving models](https://oxagen.sh/blog/pillars/self-improving-models) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [A model cannot grade its own homework](https://oxagen.sh/blog/the-limits-of-self-correction-and-model-collapse) Without outside feedback, self-correction fails, and a model trained on its own output gets worse. This post covers what the collapse and verifier research says to do instead. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/deterministic-coding-agents-every-turn-on-the-record) * [Coding agents](https://oxagen.sh/blog/pillars/coding-agents) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [Deterministic coding agents: every turn on the record](https://oxagen.sh/blog/deterministic-coding-agents-every-turn-on-the-record) A coding agent changed your code and no one can replay how. What the research says about feedback from running code, random sampling, and turns you can audit. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/self-improving-models-from-star-to-self-rewarding) * [Self-improving models](https://oxagen.sh/blog/pillars/self-improving-models) ### [From STaR to DeepSeek-R1: what self-improvement means](https://oxagen.sh/blog/self-improving-models-from-star-to-self-rewarding) A vendor says the model improves itself. This is the research behind that claim, the signal that drives each training loop, and what stops each one. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/governing-an-agent-that-rewrites-itself) * [Self-evolving agents](https://oxagen.sh/blog/pillars/self-evolving-agents) * [Code graphs](https://oxagen.sh/blog/pillars/ontologies) ### [Governing an agent that rewrites itself](https://oxagen.sh/blog/governing-an-agent-that-rewrites-itself) An agent that edits its own code needs the same review as any other change. What safety research says about reward hacking, sandboxes, oversight, and typed contracts. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/graph-grounded-retrieval-vs-vector-search) * [Code graphs](https://oxagen.sh/blog/pillars/ontologies) ### [Graph-grounded retrieval vs vector search](https://oxagen.sh/blog/graph-grounded-retrieval-vs-vector-search) Vector search finds the passage that looks like your question. Graph-grounded retrieval finds the fact that answers it. What the research says about the difference. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/self-evolving-agents-what-the-evidence-shows) * [Self-evolving agents](https://oxagen.sh/blog/pillars/self-evolving-agents) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [Self-evolving agents: what the evidence shows](https://oxagen.sh/blog/self-evolving-agents-what-the-evidence-shows) What changes when an agent improves itself, what checks the change, and the measured gain, across eight systems from Voyager to AlphaEvolve. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/the-science-of-ai-agents-from-react-to-tool-use) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [The science of AI agents: from ReAct to tool use](https://oxagen.sh/blog/the-science-of-ai-agents-from-react-to-tool-use) Planning, tool use, memory, and reflection each come from a paper that measured something. This post traces those papers and what agents still cannot do. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/what-an-ontology-buys-an-agent) * [Code graphs](https://oxagen.sh/blog/pillars/ontologies) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) ### [What an Ontology Buys an Agent](https://oxagen.sh/blog/what-an-ontology-buys-an-agent) An agent can answer with confidence from the wrong context. This post covers what classes, relations, constraints, and dated facts add to an agent's answers. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/what-swe-bench-measures-and-what-it-misses) * [Coding agents](https://oxagen.sh/blog/pillars/coding-agents) ### [What SWE-bench Measures, and What It Misses](https://oxagen.sh/blog/what-swe-bench-measures-and-what-it-misses) Coding agents are ranked by their SWE-bench resolve rate. This post covers what that rate shows, where test-based grading goes wrong, and what the rate cannot tell you. September 9, 2026 · 8 min read [](https://oxagen.sh/blog/why-agents-fail-measuring-reliability-and-cost) * [AI agents](https://oxagen.sh/blog/pillars/ai-agents) * [Coding agents](https://oxagen.sh/blog/pillars/coding-agents) ### [Why agents fail: measuring reliability and cost](https://oxagen.sh/blog/why-agents-fail-measuring-reliability-and-cost) AgentBench, WebArena, GAIA, SWE-bench, and tau-bench each measure a different thing. None of them reports what a run costs. This post covers what that hides. September 9, 2026 · 8 min read