Agent learning: why deployed agents should learn from the whole product, not just chat
techNmak · x · 2026-09-23
After months interviewing CTOs, Chief AI Officers and product leads running agents at scale, the author argues most AI strategies optimize the wrong data source: they only learn from chat. Key points:
- What agent learning is: deployed agents improving from production interactions instead of plateauing at launch quality; every conversation, correction and in-app action is a signal.
- How it works: signals become skills applied via in-context learning (no retraining needed) or curated into fine-tuning datasets.
- Three buckets: orchestration makes agents act, agent-user connectivity (e.g. AG-UI) puts them in front of people, learning determines whether agents compound in value — and it's the least built-out bucket.
Linked content is CopilotKit marketing, but the core claim — collect learning signals across the entire product surface, not just conversations — is a useful thesis for agent builders.
Related event: Product Behavior Signals Are 10-100x Richer Than Chat Logs(2 posts)→
More from coding & agent
- New Opus fixes Rust image-scanning bug with temp symlinks, impressing the developer — generativist · 2026-09-23
- Google's MCP Apps extension lets MCP servers return interactive UI components — rseroter · 2026-09-23
- Dev runs 32 coding agents burning 5-10B tokens a day in Zed — generativist · 2026-09-23
- Stanford paper: self-organizing agent team hits 66.7%, beats best single model's 48.8% — dair_ai · 2026-09-23
- Google's ToolGrad reverses tool-use data generation, hits 83.1 BFCL with 12B Gemma-3 — dl_weekly · 2026-09-23
- Using AI agents to book things still costs as much mental effort as doing it yourself — nabeelqu · 2026-09-23