11–20 of 53 posts
Best of 3 (or more) tournaments are a good strategy. You can also use them for RL via GRPO if you're running an open weight model.
Failures of workflows signal assumption violations that ultimately should percolate up to humans. Also, static dags are more amenable to human understanding than dynamic task decomposition. Robustness in production is good though, if you can bound agent behavior. Best of 3 (or more) tournaments are a good strategy. You can also use them for RL via GRPO if you're running an open weight model.
I was sort of thinking about a similar idea recently. What if you wrote something like a webserver that was given "goals" for a backend, and then told agents what the application was supposed to be and told it to use the backend for meeting them and then generate feedback based on their experience. Then have an agent collate the feedback, combined with telemetry from the server, and iterate on the code to fix it up.…
My next thought was to implement a multi agent workforce on top of this where it’s fully virtuous (like a cycle) and iterative.
https://github.com/swetjen/virtuous
If you’re interested in working on this together my personal website and contact info is in my bio.
The concept of mapping 'exceptions as observations' rather than failures is the right mental shift for production.
Question on the 'Homeostasis' metric: Does the agent persist this 'stress' state across sessions? i.e., if an agent fails a specific invoice type 5 times on Monday, does it start Tuesday with a higher verification threshold (or 'High Conscientiousness') for that specific task type? Or is it reset per run?
Starred the repo, excited to dig into the OODA implementation.
3. What, or who, is the judge of correctness (accuracy); regardless of the many solutions run in parallel. If I optimize for max accuracy how close can I get to 100% matemathically and how much would that cost?
To be fair, achieving 100% accuracy is something even humans don't do. I don't think this is about a system just asking an AI if something is right or wrong. The "judge" isn't another AI flipping a coin, it’s a code validator based on mathematical forms or pre established rules.
For example, if the agent makes a money transfer, the judge enters the database and validates that the number is exact. This is where we are merging AI intelligence with the security of traditional, "old school" code. Getting this close to 100% accuracy is already a huge deal. It’s like having three people reviewing an invoice instead of just one, it makes it much harder for an error to occur.
Regarding the cost, sure, the AI might cost a bit more because of all these extra validations. But if spending one dolar in tokens saves a company from losing five hundred dollar, due to an accounting error, the system has already paid for itself. It’s an investment, not a cost. Plus, this tighter level of control helps prevent not just errors, but also internal fraud and external irregularities. It’s a layer of oversight that pays off.
Best regards