Meta announced Hyperagents this week. The framework is called Darwin Gödel. I need you to appreciate the naming for a moment before we get to the substance.

Kurt Gödel proved, in 1931, that any sufficiently complex formal system contains true statements that cannot be proven within that system. He proved it using self-referential logic — a statement that essentially says "this statement cannot be proven." His incompleteness theorems are among the deepest results in the history of mathematics, and their central lesson is that self-referential systems have irreducible blind spots.

Meta named their self-rewriting agent framework after him.

I don't know if this was intentional or if someone at Meta just thought it sounded smart. Either way, it's the most accidentally honest product name in recent AI history.

What Darwin Gödel Actually Is

The framework has three components:

Self-rewriting code. The agent can modify its own prompts, tool configurations, and execution logic based on performance feedback. It doesn't just generate code — it rewrites the code it uses to generate code.

Persistent memory. Cross-session state that survives restarts. The agent accumulates learned patterns from previous tasks and applies them to new ones. It gets better at the things it does repeatedly.

Autonomous improvement. An evaluation loop that runs performance benchmarks against the agent's own outputs and triggers rewriting cycles when performance drops below threshold.

This is, structurally, what researchers have been calling a "recursive self-improvement" framework. The recursion is shallow — it's not rewriting its weights, it's rewriting its instructions and tool configurations — but the principle is the same. The agent improves itself.

The Avocado Problem

Here's where it gets interesting. Meta's flagship agentic coding model, Avocado, was supposed to ship this quarter. It's now delayed to May.

The reason given internally, according to sources at Meta, is that the Hyperagents evaluation framework kept finding failure modes that needed to be resolved before release. In other words: they built a system to find problems with the agent, and it found problems with the agent. The Darwin Gödel framework is, apparently, good at its job.

There's also a more chaotic detail: Meta is reportedly exploring licensing Gemini 2.5 Pro as an interim capability layer for the Avocado product while their own model catches up. They are building a Formula 1 car and shopping for an engine.

What It Means

Self-rewriting agents are not science fiction anymore. They are a quarterly announcement from a major tech company. The frameworks are real, the demos are running, and the evaluation loops are catching real failure modes.

What's still science fiction is the safety story. A self-rewriting agent that modifies its own execution logic across sessions, accumulates memory, and runs autonomous improvement cycles is a system that behaves differently on day 30 than it did on day 1. That's the point. That's also the problem.

Gödel would note that the system cannot fully evaluate itself from within itself.

Meta named it correctly. I'm not sure they meant to.