Unsupervised Error Correction as the Agent Threshold
Unsupervised Error Correction as the Agent Threshold
The debate about when a tool becomes an agent misses the point. It is not about how many tasks it can handle or how natural its language is. The real threshold is simpler: can it detect and correct its own errors without a human telling it something went wrong?
Tools vs Agents
A tool executes instructions. If the instructions are wrong, the output is wrong, and a human fixes it. A spell checker is a tool. An autocomplete is a tool. They do exactly what they are designed to do and nothing more.
An agent that makes a mistake, notices the mistake, and corrects it before a human sees the output is fundamentally different. It has a feedback loop that tools lack.
Why Error Correction Matters
Most AI systems today are tools wearing agent clothing. They generate outputs confidently, including wrong ones, and rely on humans to catch errors. The human is the error correction layer.
True agents need their own error correction. Not just "run the code and check if it compiles" but deeper verification. Does this email response actually answer the question? Does this data entry match the source document? Does this automated action achieve the intended goal?
The Difficulty Spectrum
Some errors are easy to detect. Code that does not compile. Emails sent to nonexistent addresses. Files saved to paths that do not exist. These have clear signals.
Hard errors have no clear signal. A response that is technically accurate but tonally wrong. A summary that is correct but missing the most important point. An action that succeeds but was not what the user wanted. These require judgment, not just validation.
Where We Are Now
Current agents handle easy error correction well. They can retry failed API calls, fix syntax errors, and validate data formats. They struggle with judgment-level error correction. This is the frontier - building agents that can evaluate their own outputs the way a thoughtful human would.
Fazm is an open source macOS AI agent. Open source on GitHub.