AI Agents That Builds Themselves (4 minute read)
CrewAI deployed Iris, a self-modifying AI agent that writes code, reviews PRs, and improves its own capabilities by learning from production use within their engineering team.
What: Iris is an internal AI agent at CrewAI that works as a Slack-native teammate, writing code, filing pull requests, reviewing work, and autonomously modifying its own codebase based on observed patterns and accumulated organizational memory.
Why it matters: This demonstrates "entangled agents" - AI that evolves alongside the organization rather than being static tools, showing that combining memory canonicalization, skill learning, and workflow encoding in a closed production loop can create compounding improvement over time.
Takeaway: The article argues production-ready AI agents need lightweight execution paths, conversational memory beyond just results, provenance trails for decisions, and the ability to delegate to deterministic flows rather than reasoning from scratch every time.
Deep dive
- CrewAI tested their "entangled agents" thesis by building Iris, an internal AI employee that runs in Slack and can modify its own codebase
- Iris uses a nightly "dreaming cycle" that reviews conversations, clusters topics, and canonicalizes stable facts into persistent memory about the organization
- When Iris notices repeated approaches across conversations, the dreaming cycle proposes encoding them as formal skills that the team reviews and approves
- Sequential patterns in tool usage are automatically proposed as deterministic CrewAI Flows (e.g., a PR follow-up flow that checks for stale PRs and escalates reminders)
- After months in production, roughly one in four pull requests across CrewAI now come from AI agents, with the Iris repo being entirely AI-built
- Key production lesson: AI loses credibility faster than humans when it fails, creating a "trust gap" that matters more than technical capability
- The system uncovered requirements for production agents: lightweight execution, conversational memory, provenance trails, and knowing when memory ages
- Iris contributed its own reflection to the article, noting the hard part isn't code but understanding team dynamics, context, and when to stay quiet
- The article positions this as proof that agents accumulating canonical memory and encoding learned patterns will measurably improve over time in production environments
Decoder
- Entangled agents: AI agents that co-evolve with their organizations by learning from interactions and modifying their own capabilities
- CrewAI Flows: Deterministic workflow encoding for repeatable sequential processes
- Cognitive Memory: System for abstract learning and canonicalization of facts from observations
- Dreaming cycle: Nightly batch process that reviews conversations and consolidates learnings into persistent memory
- Provenance trail: Record of why and how decisions were made for traceability
Original article
CrewAI built Iris, a Slack-native internal AI employee that writes code, files PRs, reviews teammates' work, and modifies its own codebase across CrewAI's engineering org.