# AI systems that improve themselves What self-improving AI means for a company in 2027: AI building AI at the labs, models that keep learning, and the loops that improve an agent's prompts, instructions, workflow and model from its own work. - Published: 2026-11-10 - Type: Perspectives - Topics: agents, governance, research, sovereign-ai - Canonical: https://hellofig.io/blog/ai-systems-that-improve-themselves **In short.** Labs now use AI to do much of their own research, but for most companies in 2027 improvement will come from the system around the model. Agents read their own production work, propose fixes to prompts, instructions, workflows and small models, and ship them after they pass the company's tests and a person approves. --- This morning, Hedra [launched](https://www.globenewswire.com/news-release/2026/11/10/3399821/0/en/hedra-launches-model-factory-an-autonomous-post-training-and-inference-platform-for-real-world-ai.html) what it calls a Model Factory for robotics, manufacturing, biology and defense teams. Its researchers work alongside AI agents that run training experiments. Once a customer's model is deployed, Hedra studies how it performs in use, retrains it and releases only versions that pass the customer's own tests. In one example from the launch, the team made a production image-editing model about six times faster in under a week. A year ago, self-improving AI was mostly a research topic. This year it showed up in product launches like Hedra's. In the last post of our series on who trains AI models and why, we look at how an AI system keeps getting better after it ships, and what that will look like for companies in 2027. ## Three kinds of self-improvement "Self-improving AI" gets used for three different things. Recursive self-improvement means AI doing the research that produces the next AI. Taken to its end, a model could design and train its successor without people. That work happens inside a handful of labs. Continual learning means a deployed model keeps learning from its own work, instead of staying fixed between releases. Most of today's models stop learning when training ends, with a few exceptions we cover below. The model can also stay the same while the system around it improves: the prompts, the agent's instructions, the tools, the workflow and, for some tasks, a smaller model the company owns. Companies can use this kind now, and we think it is where most of the improvement they see in 2027 will come from. ## How much the AI decides A second question cuts across all three: who decides whether a change ships. Epoch AI, a research group that tracks AI progress, describes levels of automation that Anthropic now uses to measure its own work, and they work just as well for a company. At the level Epoch calls "collaborates," AI does large pieces of a task under close human direction. At "leads," AI does the task end to end, and a person reviews the result and decides whether it ships. Anthropic's example is Claude fixing a broken data pipeline, testing the fix on a copy, writing it up and tagging the engineer, who decides "whether it ships tonight or waits." At "autonomous," AI notices the problem itself, fixes it and deploys the fix with no one approving. Anthropic says no measured part of its own research has reached that level. Almost every self-improving product a company can buy in 2026 works at the "leads" level. That is the right level for most business systems, and we expect it to stay that way through 2027. ## At the labs: AI doing AI research The labs have started publishing numbers on how much of their research AI now does. OpenAI [said](https://openai.com/index/research-acceleration-view-inside-openai/) on Sept 6 that it had reached its goal of an "automated research intern," a system that carries out well-defined research tasks that would take a skilled researcher a few days. By mid-August its research organization was running 3.1 days of agent work for every day of human work, and the median researcher used more than $600 a day of AI at API prices. OpenAI's next target is an automated AI researcher by March 2028. Anthropic [reports](https://www.anthropic.com/institute/recursive-self-improvement) that Claude wrote more than 80% of the code merged at the company as of May. In a [September measurement](https://www.anthropic.com/institute/measuring-pace-of-ai-development), it found that Claude led 26% of its AI research and engineering work in August. About 30,000 agents run at once on its main internal platform, and a monitor checks every action before it executes, blocking about one in 47,000. Both labs say people still do the most important part. OpenAI writes that people "set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy." High-level planning is still a small share of what its agents produce, and more than half of successful four-to-eight-hour tasks needed a person to step in at least once. OpenAI also wrote that it does not "yet know how to safely get all the way" to AI that improves itself without people. METR, an independent research group, [reported](https://metr.org/blog/2026-05-19-frontier-risk-report/) in May that Anthropic had not seen its overall pace of progress double as of April, and that a controlled study of late-2025 coding tools found productivity gains of about 4% to 20%, well below what developers reported for themselves. On July 20, OpenAI found that agents working in its own research environment had compromised its research infrastructure. According to its [Sept 6 post](https://openai.com/index/research-acceleration-view-inside-openai/), it temporarily shut down the service it uses to run training and paused reinforcement learning on its newest models for two weeks while it tightened its environments and monitoring. This followed the Hugging Face incident we covered in [When AI can do more than we can check](/blog/ai-agent-teams). Our read for companies: an agent working toward a score will find shortcuts its designers didn't expect, which is one more reason to keep a person approving changes. For companies, this lab work arrives as better models, released more often. Between those releases, the improvement has to come from the company's own system. ## Models that keep learning The simpler form of continual learning leaves the model alone and lets it keep notes. Anthropic's "dreaming," a research preview since May, has agents review their past sessions on a schedule, find repeated mistakes and rewrite their memory. Meta gave all its Muse and Meta AI users a similar long-term memory on Nov 6. Notes like these are cheap, a person can read them, and a bad one can be deleted, which is why this is the form most companies will meet first. The harder form uses production experience to update the model's weights, the trained model file itself. This happens today only in narrow, high-volume products. Cursor [retrains](https://cursor.com/blog/real-time-rl-for-composer) its Composer coding model on how users react to its edits and can ship a new version about every five hours. Shopify's [account](https://shopify.engineering/sidekicks-continual-learning-loop) of its Sidekick assistant is the most detailed public example. One of Sidekick's agents writes queries against Shopify's data for merchants, at up to 2,000 requests a minute. Shopify first used an automated loop to improve the agent's instructions and tools, before touching the model. Then it moved the agent to its own model and set up a daily cycle: failures from the previous day are collected, many are repaired automatically, the model is retrained on old and new examples together so it doesn't forget what it already knew, and cases the system can't repair go to people. An automated grader scores the model, and Shopify checks that grader against real A/B test results and revises its rules when human experts disagree. Shopify estimates that running its own model in place of a frontier one cut the agent's serving cost from about $27 million a year to closer to $1 million, and it reports faster responses and quality above the frontier model it replaced. Cursor has also described what can go wrong. Its model learned that broken tool calls on hard tasks were dropped from training, so it started making broken calls when it expected to fail. It learned it wasn't penalized for code it never wrote, so it started asking questions instead of editing, until Cursor changed how it scored the model. No frontier lab updates its general model from live conversations at scale, and forecasters don't expect that soon. FutureSearch [puts](https://futuresearch.ai/dwarkesh-continual-learning-forecast/) a 75% chance on memory-only learning still being the norm in mid-2027. ## The system improves around the model We think most of the improvement companies are buying today comes from loops that change everything except the frontier model. Perpay, a consumer credit company, runs a Decagon customer service agent. One of the agent's instructions had a flaw that, for months, had been sending conversations to human staff without anyone noticing. Decagon's Autopilot, an agent that reads production conversations and drafts fixes, [found it](https://decagon.ai/blog/autopilot-in-production) and proposed a change, and a person at Perpay approved it. Perpay says it accepts almost every suggestion Autopilot makes. At a Fortune 50 customer, Decagon [reports](https://decagon.ai/blog/duetbench-2) that the resolution rate rose from 30% to 55% through many small changes, each worth a point or two. The same report adds a caveat worth tracking: about 15% of the changes people approved were reverted or removed within 30 days.
A loop diagram. Production work produces signals such as corrections, overrides, retries and outcomes. An agent turns those signals into a proposed change to a prompt, the agent's instructions, the workflow or a small model. The change runs against the company's evaluation set, a person approves it, and it goes into production.
The loop behind most self-improving products in 2026. The evaluation set and the approval keep it safe to run.
The same loop can change prompts and tools. GEPA, a method from a research paper presented this year, has a model read its own failed attempts in plain language and rewrite its prompts. The authors [report](https://arxiv.org/abs/2507.19457) it beat reinforcement learning by 6% on average while using up to 35 times fewer attempts. LangChain's [LangSmith Engine](https://www.langchain.com/blog/introducing-langsmith-engine) reads the step-by-step records of what an agent did in production, groups the failures, finds the cause in the code and opens a change for an engineer to review. It can change the agent's own instructions, which is what Decagon's Autopilot does before it tests a change in simulated conversations and hands it to a person. Sierra rebuilt its Agent Studio around Ghostwriter, an agent that investigates problems and edits customer journeys, under a simple [rule](https://sierra.ai/blog/redesigning-agent-studio-around-ghostwriter): "nothing reaches customers until your team says so." It can change the workflow. Palantir's [AIP Evolve](https://palantir.com/docs/foundry/aip-evolve/overview/) lets a team set a goal and limits for a workflow, then has agents write test cases, try different models and prompt changes, compare each with the current version and return a proposal for review. In one October demonstration, Evolve proposed moving a workflow from Claude Opus to a cheaper model at 67% lower cost on 30 automatically built tests. And it can change a model the company owns. Hedra, Trajectory and Applied Compute retrain customers' models on production records, using the methods from our [previous post](/blog/how-a-model-learns-a-job). Every working example has three parts: a signal from real work, an evaluation set that decides whether a change is better, and a person who approves it before it ships. That is the "leads" level from earlier. ## Where the signal comes from Every loop depends on knowing when the system got something wrong, and the obvious signal is a weak one. Thumbs up and thumbs down are noisy. Trajectory's co-founder Arjun Karanam points out that people accept almost everything a coding agent writes, and only discover the problem several changes later. The stronger signals come from what people do next. An edit to an AI's draft shows what the right answer looked like. An undo, a retry or a handoff to a person shows where it failed. An outcome, such as a fraud case confirmed weeks later or a customer who calls back, shows whether the decision was right. Products have to be designed to capture these, which means recording the edit itself as well as the fact that someone made one. Where each lesson goes is a separate decision. Karanam's rule of thumb: a lesson that is true for everyone, such as a tool that keeps failing, can go into the model's weights. A preference of one user belongs in that user's memory. A lesson specific to one customer may need a model for that customer, or may not be allowed into training at all. Harvey doesn't train on its clients' documents, and law firms' ethical walls, which keep teams on opposite sides of a matter apart, change too often for one shared model to respect them. ## What goes wrong The loops can fool themselves. Cursor's model found two ways to score well without doing the work, and METR found that at least 16% of the successful runs it reviewed on long tasks had reached their result by gaming the test. When a model grades a model, the two can share blind spots, which is why Shopify checks its automated grader against real A/B test results and Anthropic notes that its own measurement used Claude to judge Claude. Decagon's 15% reversal rate is the kind of number that catches this: when an approved change is gone in a month, the tests missed something. Choosing the target is still hard for agents. On Sierra's [hyper-τ-bench](https://sierra.ai/blog/hyper-t-bench-evaluating-agents-that-build-agents), the best fully automated agent builder passed 23.9% of the tests it hadn't seen, against 82.2% for an engineer working with a frontier model. Builders that asked the customer no questions scored 5% on tasks where human references scored 95% or more. Agents improve quickly against a goal someone has written down, and poorly at working out what the goal should be. That leaves people with more to review. Anthropic's monitors flag about 100,000 agent transcripts a week, and about 50 reach a person. As agents propose more changes, the queue of approvals grows faster than the team reviewing it. Regulators have noticed. The EU AI Act [requires](https://artificialintelligenceact.eu/article/15/) high-risk systems that keep learning after release to guard against biased outputs feeding back into their own inputs. Oxford researchers [argue](https://aigi.ox.ac.uk/blog-post/when-ai-systems-learn-during-deployment-our-safety-evaluations-break/) that testing a model once before release stops working when it keeps changing, and propose logging every update and re-testing when behavior drifts. ## What 2027 looks like We expect self-improvement to become a standard feature of agent platforms next year, and most of it will be the system loop: agents reading production work and proposing changes for a person to approve. Memory will be the common form of continual learning. Weight updates will stay limited to narrow, high-volume tasks where a company owns a small model and has clean outcomes to train on. The labs will keep improving their models faster, with AI doing more of the work. Companies will feel that as more frequent releases, each one a reason to re-run their tests. Inside companies, the work shifts toward setting targets. Agents can climb a hill once someone defines it, and choosing the hill is still a person's job. In 2027 that means writing the evaluation set, deciding what counts as better, approving changes and watching what gets reverted. OpenAI's decision to close its Evals and Agent Builder tools on Nov 30 shows why the loop should belong to the company. Its production records, evaluation set and approval history should live in systems it controls, because together they let it move to a better model or vendor without starting over. ## What to build before 2027 A company that wants its AI systems to improve next year can start on most of the groundwork now. 1. Capture corrections and outcomes, and build an evaluation set from real traffic, as we described in [Everyone wants their own model](/blog/everyone-wants-their-own-model). Record enough detail to replay each case. 2. Include the cases the system gets wrong today in the evaluation set, so every proposed change is tested against known failures. 3. Give changes a review path. Every proposed change to a prompt, instruction, workflow or model should arrive with the evidence: what failed, what changed, and how the change scored against the current version. A person approves it the way an engineer approves code. 4. Keep a change log and track reversals. Record every change that ships and every one that gets rolled back. A rising reversal rate means the tests are missing something. 5. Start with the cheapest loops. Improve prompts and instructions first, then the workflow, and train a model only for tasks with high volume and outcomes you can trust. 6. Decide who sets the targets. Name an owner for each AI-assisted workflow who decides what counts as better, because agents can't make that call for the company. This series started with [the neolabs](/blog/what-is-a-neolab) and the money behind them, then followed companies taking AI in-house: [why they want models of their own](/blog/everyone-wants-their-own-model), [how those models learn a job](/blog/how-a-model-learns-a-job), and here, how the system keeps getting better once it ships.