Signals: Toward a Self-Improving Agent
Source: https://factory.ai/news/factory-signals Author: Factory.ai Date: 2026-01-27
Summary
Factory.ai (AI software development platform) announces Signals — a system that watches developer sessions, identifies friction points, and automatically improves the agent’s capabilities to address them. First real-world “self-improving agent” loop in a product context.
Key Claims
- Signals observes: when users undo agent changes, add manual corrections, or abandon agent suggestions — these are implicit negative feedback signals.
- Signals identifies: patterns of failure (recurring correction types, specific code patterns that always need fixing).
- Signals improves: updates the agent’s system prompt, tool configuration, and few-shot examples based on failure patterns.
- No human labeling: the improvement loop runs fully automatically from user behavior.
- Key limit: Signals doesn’t retrain the model — it adapts the harness. Model weights are static; context and configuration adapt.
- This is the Agent Labs thesis (swyx) made concrete — operational data → system improvement.
Entities
- Factory.ai — AI software development platform
Concepts
- Coding Agents — Factory’s product domain
- Autonomous Research — self-improving loop as autonomous improvement
- Context Engineering — system prompt adaptation as context engineering
- RL Infrastructure — implicit RL from user behavior