A self-improving AI agent

Iris is CrewAI's internal AI employee, working in Slack to write actual code, create and review PRs, onboard tools, and even modify its own code. Running in production alongside the engineering team for several months, CrewAI used it as a testbed for its entangled agents hypothesis.

The core idea is to remember repeated behaviors and turn their outcomes into more refined units of execution. A nightly dreaming cycle groups conversations by topic and organizes stable facts into canonical memory, building a structure that approaches a World Model that understands the whole company.

• Memory understands conversations and remembers stable facts, such as trusted credential sources or team-specific repositories.
• Skills turn repeated non-deterministic workflows into formal skills, which are added after human review and approval.
• CrewAI Flows lock in tasks that repeat in the same order every time, like PR follow-up actions, as deterministic flows.

Operations weren't smooth. Sandbox connections dropped, some proposals were rejected, and it was confirmed that AI loses trust much faster than humans do. At the same time, production principles became clearer, such as lightweight execution paths, per-conversation memory, awareness of memory aging, and a provenance trail.

The results were clear too. The AI agent has merged hundreds of PRs across CrewAI's repositories, accounting for about a quarter of all PRs across the organization. However, success isn't measured by PR count alone, and more output doesn't automatically mean better outcomes. The Iris repository was written entirely by AI, and a division of labor has taken hold where humans focus on architecture, design, and judgment while the agent handles the volume of implementation.

The conclusion is simple: it's not the agent with the flashiest demo that survives, but the one that holds up in production and builds trust over time.