"ACON" may be the most worthwhile AI term you learn this month. It targets an annoying problem: AI keeps making the same mistake. The usual fix is to hand-edit the prompt, only for it to fail somewhere else. ACON takes a different approach: every time the AI fails, it automatically distils a "guideline" and stores it, then applies it the next time a similar situation arises, making fewer mistakes over time. This article unpacks what ACON is, how the mechanism works, and what to watch when adopting it in an enterprise.
What is ACON?
ACON is a technique proposed in 2026, best understood as "failure-driven guideline optimization" for agents. Its core is to let an AI agent learn from its own failures: whenever a task goes wrong, the system analyses the cause and distils a natural-language "guideline," storing it in a guideline library for future similar tasks to read and apply. The goal is to turn a one-off mistake into experience that can be reused over the long term.
The problem it solves: AI repeating mistakes
Many people using AI agents hit the same pain point: it repeats mistakes in similar spots; you correct it once, and next time in a new context it errs again. The root cause is that an ordinary agent does not record "why it went wrong this time," and so has no way to avoid a repeat. ACON fills exactly this gap, giving the agent the ability to learn from failure instead of starting from scratch each time.
The core mechanism: failure-driven guideline optimization
ACON works as a closed loop: the agent runs a task, hits a failure, analyses the failure, writes a guideline, and applies it next time. When a run falls short, the system reviews that trajectory, pinpoints the key misstep, and condenses "what to do next time in this situation" into a clear guideline. That guideline is then added to the agent's guideline library, becoming a reference for the next decision.
How mistakes become reusable guidelines
The essence of ACON is turning mistakes into assets. Each failure, after analysis, settles into a natural-language, human-readable guideline, for example "when handling this type of invoice, verify the date format before posting." These guidelines are independent of any single conversation, stored long-term, and retrieved and applied at the right moment. In other words, every mistake the agent has made becomes a rule that keeps it from repeating it.
Why it grows more stable with use
Because the guideline library accumulates with use, the more failures the agent has seen, the richer the guidelines it can follow, and the lower the chance of repeating them. This is a virtuous cycle: more use builds more experience, more experience means fewer mistakes, and fewer mistakes mean steadier performance. For long-running, highly repetitive workflows, this "steadier with use" property is exactly what reliable automation needs.
How it differs from hand-editing prompts
The traditional way is: the AI errs, you hand-edit the prompt, then wait for the next error. This is labour-intensive and hard to cover every case. ACON automates the loop: the agent analyses its own failure, writes its own guideline, and applies it itself. The human role shifts from "fighting fires one by one" to "reviewing and gatekeeping" the quality of the guideline library, sharply improving both efficiency and coverage.
Best use cases
ACON delivers the most value in long-running, highly repetitive workflows where mistakes follow a pattern: automated customer service, data processing, report generation, cross-system operations. These repeatedly meet similar situations, letting the guideline library compound its "learn once, benefit long-term" effect. Conversely, if a task is extremely one-off and pattern-less, the value of accumulating guidelines is limited.
Adoption considerations and common mistakes
Watch the management of the guideline library: the more guidelines, the more you must prevent them from conflicting or going stale, requiring periodic review and pruning, or they will drag down judgment. The three common mistakes are: first, letting guidelines accumulate endlessly without curation, causing rules to clash; second, writing guidelines too vaguely, leaving the AI nothing concrete to follow; third, treating ACON as an excuse to remove human oversight, ignoring review of high-risk guidelines. Govern the guideline library well, and ACON truly makes an agent more reliable with use. Want to know how ACON lands in your workflow? Visit ai.ud.hk to explore UD's AI Staff solutions and see how to make AI improve continuously from its failures.
懂AI，更懂你｜UD相伴，AI不冷
