September 2026 Summaries

Self-improving AI agents aim to reduce the manual effort of reviewing failures and rewriting prompts by using continuous evaluation and automated iteration. The proposed teacher-student framework compares a trusted, often expensive agent with a cheaper candidate agent on the same tasks, while an independent evaluator scores their outputs and a coach agent uses feedback to suggest revisions, such as prompt changes. Changes are repeatedly tested across successful and failed cases to identify regressions, with human reviewers approving any candidate for deployment and retaining the original agent as a fallback. A support-ticket example demonstrates how instructions can evolve to correct policy reasoning errors while preserving requirements such as response length. The approach resembles optimization through “textual gradients,” where natural-language feedback guides changes but test results determine whether they are retained. Its limitations include flawed evaluators, overfitting to familiar cases, incorrect teacher standards, degrading performance across iterations, operational costs, and problems that prompts alone cannot solve, requiring safeguards, fresh evaluation data, controlled deployment methods, and human oversight.

Komodor announced the Agentic Operations Platform, a system designed to help SRE, DevOps, and platform teams deploy, govern, and optimize autonomous operational workflows for incident management, reliability, software operations, and cost reduction. The platform combines more than 50 configurable pre-built agents, integrations, and workflows with tools for organizations to create, import, and manage custom agents using shared context, persistent memory, knowledge graphs, governance controls, and support for on-premises or cloud deployments. It addresses concerns that agentic AI projects may fail because of weak governance, unclear returns, and uncontrolled costs through role-based access policies, behavior guardrails, human approval for risky actions, spending limits, auditing, scenario-based evaluation, and shadow testing. Komodor says the platform builds on five years of experience with its AI SRE offering and is available immediately through the company and its worldwide business partners.