Self-improving AI agents are systems that rewrite parts of their own set-up, such as their instructions, tools and memory, after studying where they failed. The research that works keeps one thing out of the agent’s reach: the test it is judged by. For a business, that test should be built from its own past work and owned by a named person.
Agents that tune themselves are moving out of research papers and into the products businesses buy. That is useful, and it changes the question a leadership team has to answer: from whether an agent is good to what it is good against, and who decided.
What are self-improving AI agents?
An AI agent is a model wrapped in the engineering around it: the instructions it follows, the tools it can call, what it remembers between tasks and the checks on its work. A self-improving agent changes that wrapping itself. It runs a batch of tasks, reads the record of what went wrong, proposes an edit to its own instructions or tools, and keeps the edit if the results improve. The model underneath usually stays the same; the system around it learns.
That wrapping matters more than most buyers assume. In our experience, the model accounts for about 20% of how well an agent performs; the engineering around it, meaning how the working loop is designed, how it connects to your systems and how it is tested against your own past cases, is the other 80%. A self-improving agent automates part of that 80%, so the question of who checks it moves with it.
Is a self-improving AI possible?
Yes, within limits, and the published results are real. The Darwin Gödel Machine, a system that repeatedly edits its own code and tests each change on coding benchmarks, raised its score on SWE-bench from 20.0% to 50.0% and on Polyglot from 14.2% to 30.7% (Zhang et al., arXiv 2505.22954, revised March 2026). In Agentic Harness Engineering, ten rounds of self-edits lifted a coding agent’s pass rate on Terminal-Bench 2 from 69.7% to 77.0%, ahead of a human-designed set-up at 71.9% (Lin et al., arXiv 2604.25850, April 2026).
The limits matter as much as the gains. Those results come from software benchmarks, where a test passes or fails in seconds. The Darwin Gödel Machine’s authors ran every experiment with sandboxing and human oversight. And not every model gains: a May 2026 study across model sizes found that the weakest models benefit little from an improved set-up, because they fail to use it or to follow it faithfully (Lin et al., arXiv 2605.30621). Most business work has no instant pass-or-fail test.
What stays fixed while the agent changes itself
The systems that improved reliably share one design rule: the agent can change its own working parts, never the thing that grades them. In Agentic Harness Engineering, the agent may write only inside its own workspace; the run records, the tracing, the verifier and the model settings are read-only, which the authors say blocks shortcuts such as switching off the verifier or swapping the model. Each edit also carries a written prediction of what it will fix, which the next round confirms or reverts. In Self-Harness, an edit is kept only if it improves results without measurable harm on tasks held back from the agent, and rejected edits are logged without touching the live set-up (Zhang et al., arXiv 2606.09498, June 2026).
In a business, the same rule sorts every part of an agent into one of two columns, each with an owner:
| The agent may change | Stays fixed | Owned by |
|---|---|---|
| Its instructions | The past cases it is judged against | The process owner |
| Its tools and how it calls them | What it may read and write | Whoever is accountable for the data |
| What it remembers between tasks | The rule for accepting a change: better on the cases, worse on none | A named person who signs off |
| Its working notes | The log of every change, kept or rejected | The team that runs it |
This is how we build AI already: tested, back-tested, versioned, released in stages, and reversible. An agent that proposes its own changes runs that discipline more often, so the fixed column has to exist before the agent starts.
AI agent evaluation starts with your own work
A public benchmark tells you whether an agent can fix code in someone else’s software. It cannot tell you whether it handles your credit notes, your complaints or your exceptions the way your best people do. For that you need a test set built from your own past cases, with the answers your senior people would give.
Agents are also poor judges of their own success. In four attempts to have a chain of AI agents write research papers end to end, three failed during implementation or evaluation, and one of the recurring failures was declaring success despite obvious failures (Trehan and Chopra, arXiv 2601.03315, January 2026).
Building that test set is the Capture it step of our method. We map how the work moves in practice (See it), delete what should not exist (Simplify it), then write senior judgement down as rules, decision trees and worked cases (Capture it) before anything is automated. That captured judgement is what an agent is measured against, whether a person or the agent itself proposes the next change.
Where we start
We fund research as well as reading it, including a four-year PhD with the University of Derby on federated learning, and we read papers like these to decide what is ready for a client’s live work. Self-improvement is ready for narrow jobs with a clear test. For anything else, deciding what good looks like comes first.
When the work is chosen well, the return is measurable. Our work with one manufacturer took its weekly planning run from 8 hours to under a second.
Our earlier article on the mistakes we see businesses make with AI agents covers the named owner, the audit trail and the off switch. Digital Planning is a board-level technology partner. We diagnose how a business runs, prove the change on the client’s own systems, then install, govern and run it. If you are weighing an agent that tunes itself, talk to us about the yardstick first.



