Rethinking Continual Learning: Learn the Pattern, Not the Correction
When an agent gets corrected, the question is not just how to update it. It is what the correction teaches us about the next hundred cases.
What Should AI Agents Learn From?
Our previous post showed that AI agents can improve after long-horizon work. That is only half the problem. Once a trajectory produces an outcome, a review, or an error, the system still has to decide what becomes durable know-how.
It is tempting to say: learn from everything. In production, that breaks down. Good teams do not turn every review note into policy. They ask whether a note is a one-off, a noisy opinion, or a sign of something deeper. Perpetual treats continual learning the same way: find the reusable failure pattern inside noisy work, validate it, and update the right layer of the agent.
Why the Signal Is Hard
In long-horizon work, supervision is often indirect. A reviewer may comment on the whole run, point to one bad step, approve a decision, or simply reveal that the outcome failed. The real world has no evals, only outcomes. The basic loop captures this after-the-fact signal and tries to learn from it.
A simpler agentic learning loop trains directly on that signal. This is useful, but it treats the correction as the lesson. The harder question is whether the same weakness appears elsewhere.
Why Every Correction Is Not Enough
Imagine an employee reviewing an agent that ranks refinance opportunities. The review says:
"This borrower locked a very low fixed rate during the COVID-era rate cycle. Home equity alone is not enough to infer refinance intent."
There is probably something useful in that review. But one review is not enough to change an agent. Before anything becomes durable, three questions matter:
- Is the signal reliable? Feedback may come from another language model, a hurried reviewer, an incomplete tool trace, or a preference that is not objectively correct.
- What lesson generalizes? The lesson is not about one borrower. It is a broader hypothesis: borrowers with unusually low existing rates may be structurally less likely to refinance unless the economic incentive is verified.
- Does the weakness recur? One task may expose a weak refinance recommendation; another may show the agent missing economic context; a third may show it trusting a surface-level signal. The recurring weakness is more valuable than any single correction.
From Correction to Auto-Discover
So the question changes. Instead of asking
How do we learn from this correction?
we should first ask
What hypothesis does this signal support?
That question matters because learning is compression. People do not keep a raw tape of every mistake. The brain filters noise, consolidates facts, turns repeated actions into habits, and uses error signals to ask why an action failed. We do not remember every correction as a separate rule. We learn the pattern that would have prevented the mistake.
Auto-Discover gives agents that missing pause before they learn. It looks at a trajectory and asks: is this a fact, a behavior, a failure mode, or just noise? Promising explanations become candidate hypotheses. Auto-Discover checks them against related examples, drops the ones that do not hold up, and sends the validated pattern to the right place: memory, the agent harness, or weights.
The contrast is simple. A direct loop captures feedback and updates immediately:
Perpetual inserts Auto-Discover before the update:
That routing decision is the heart of the system: what to keep, what to ignore, and which layer of the agent should change.
Validated Patterns as the Unit of Learning
After enough work, the signal starts to separate. Across many trajectories, Auto-Discover can distinguish reusable facts, persistent behaviors, repairable failures, and noise. It surfaces patterns such as:
- over-weighting surface-level signals while missing economic context,
- over-trusting retrieved documents without checking whether they apply,
- selecting the wrong tool,
- failing to ask clarifying questions,
- committing before gathering sufficient evidence.
Once Auto-Discover validates a pattern, it becomes a reusable unit of improvement. One corrected trajectory can seed related tasks, targeted hindsight feedback, and evaluations that check whether the weakness is actually fixed. The result might be a memory entry, a harness change, or a weight update. The key is that the system learns a generalizable pattern, not a private example.
The Perpetual Loop
Together, the two posts describe Perpetual's loop: learn from real work, use Auto-Discover to test what generalizes, and keep only what should become durable know-how. The application is now enterprise agents, but the principle comes from our automated red-teaming work: robust systems improve when they find the mechanism behind a failure. Patterns transfer. Data doesn't.
Today’s AI agents perform work. Perpetual agents learn from the work, and Auto-Discover helps them keep what matters.