open engineering problems

[honganh1206]

• How to balance unpredictability against structured programming without killing the agent’s ability to problem-solve; in other words, titrating in just the right amount of nondeterminism.
• How best to connect agents to ground truth so they can’t lie to themselves about having solved a problem to early-exit their loops.
• How to connect agents (which, again, are really just arrays of strings with a JSON configuration blob tacked on) to do multi-stage operation, and what the most reliable intermediate forms are (JSON blobs? SQL databases? Markdown summaries) for interchange between them
• How to allocate tokens and contain costs.