First, let’s be honest: what you read here is a vision, not a capability we have today. This is a forward-looking roadmap — a direction we believe in, but haven’t yet walked. We want to talk about an idea that matters to us: an agent that learns from its own work.
Today: starting from scratch
Today, an intelligent agent begins each conversation almost anew. It may have access to memory, but it doesn’t learn from the quality of its own work yesterday. If one of its proposals was rejected, that rejection doesn’t become a lasting lesson. If there’s a pattern in its mistakes, it doesn’t see it itself. The agent works, but it doesn’t reflect on its own work.
The vision: meta-cognition
The idea we’re thinking about is adding a layer of meta-cognition: the ability to think about its own work. In this vision, the agent has access to a lasting picture of its own behaviour — facts about how it has acted, notes from human-reviewer corrections, and patterns in the history of its accepted and rejected proposals. The agent reads this picture and gradually adjusts its behaviour. Each human correction, instead of being forgotten, becomes a lesson.
Why we haven’t built it yet
You might ask: if this idea is valuable, why isn’t it here today? The honest answer is that we’ve deliberately deferred it. At the current stage, the complexity and cost of this layer aren’t justified by the benefit it brings. Our priority is building solid foundations — a system that is reliable and understandable today. Meta-cognition is a layer that builds on those foundations, not before them. This is a conscious decision, not an oversight.
Why we talk about it anyway
If we haven’t built it, why mention it now? Because we believe direction matters as much as the step. We want to be clear about what we’re building and what we aspire to — and not confuse the two. The self-learning colleague is an aspiration we believe in and leave room for in today’s design, but we don’t sell it as today’s reality.
Putting it together
The difference between a tool and a colleague may be just this: a tool is the same each time, but a colleague gets better with time. We believe in this vision — an agent that learns from its own work — and we step toward it. But until the day it arrives, we’ll call it what it is: a forward-looking roadmap, not a promise for today.