Agentic Self-Improvement Is Becoming Its Own Field

A survey of self-improvement in modern agentic systems maps how agents improve prompts, memory, tools, plans, and workflows, while showing why guardrails matter.

1 min1 source

In one minute

  • A survey of self-improvement in modern agentic systems maps how agents improve prompts, memory, tools, plans, and workflows, while showing why guardrails matter.
  • The record is connected to 3 topics: agentic-systems, self-improvement, research.
  • 1 public source carries the evidence boundary.

Source ledger

Publishable sources attached to this record.

1 public source
#SourceRolePublic status
1arxiv.orgpaperprimary receiptsource_urls
On this page

The survey of self-improvement in modern agentic systems marks an important shift. Agents are no longer discussed only as task executors. They are becoming systems that may improve their own prompts, memory, plans, tool use, and workflows.

That is a natural next step. If an agent can inspect traces and failures, it can suggest a better rule, a new subtask, a different search strategy, or a memory update. This is also where uncontrolled loops become dangerous.

The boundary

Useful self-improvement:

  • works from traces;
  • separates suggestion from apply;
  • requires human approval for write actions;
  • is checked by tests or evals;
  • leaves a reversible change.

Dangerous self-improvement starts when the agent changes the rules that later judge its own behavior.

New Runtime Read

A self-improving agent without an audit trail is not progress. It is loss of control. The value starts when the improvement loop is measurable, reversible, and tied to external verification.