I agree. When they exist, they should be used.
I agree. And, it may never be. They’re intentionally indeterministic by design.
Agents, on the other hand, live by a constitution and increasingly, those constitutions are designed to overcome their lack of determinism.
If you examine the traces and reasoning of LLMs you see lots of assumptions that are obvious missteps. Agentic traces, depending on the agent’s programming, are very different. Here’s an actual example:
Last night I asked an agent to recommend an approach in this community question. While examining the agent’s reasoning and solution-validation steps, I saw this error.
Note: The agent often conflates Coda’s MCP with the API. This work is being performed with Coda MCP because it is a superset of the Coda API.
As you can see, there are two important distinctions we can make based on this examination:
- The agent recognized it had made an error.
- It corrected the error.
This agent, unlike general conversational LLMs running inside ChatGPT, Gemini, and Claude, operate within a set of constitutional laws. The reasoning traces call out it’s exact understanding of the requirements to verify, test with actual documents and components, and directly eval the tests against outputs and documentation. Many of these rigorous guidelines are in mcpOS, a framework that extends Coda’s MCP in ways that nudge it closer to the determinism we seek.
No LLM or agent will ever be 100% perfect. But, increasingly, they are performing in ways that are more perfect than people. I think they’re able to do this because we’re giving them more guidance in frameworks that perform deep assessments of their own work, which are processes that humans mostly skip because of constraints or are too lazy to check their work from different perspectives and scenarios.
Another significant reason these tools are different is because of the fundamental agentic loops. LLMs don’t operate in a loop based on goals, a plan, and testable outcomes. Some models are being built to contain loops within, but these are currently limited to the expensive variety and their outputs show far fewer hallucinations as would be expected.