Today my two passions collide. AI deep research and building AI agents in Tines.
If you are building agents in Tines you may be facing this situation. Your AI Agent solves a real problem mid-run, a tool needs a retry, a source is timing out, a key has to be reshaped, and then the run ends and all of that is gone. Next run, it figures it out again from zero. You notice when you watch your local coding agent. In production, unless you build it, nobody is watching the run mid-flight, so the same forgetting happens silently, thousands of times.
So I put a research question to it: can a pipeline of stateless LLM agents improve across runs, on a platform with no agent memory, using only externalized deterministic steps and a human gate?
The attached brief lays out a method I’m piloting for adding cross-run learning anyway, using only what Stories already has. Deterministic actions on either side of a stateless agent: a pre-hook that retrieves and splices advisory notes into the prompt from Records, a Resource or a skill, a post-hook that extracts what happened from steps[] and admits it to quarantine, a human gate that promotes, and dated supersession so contradictions never pile up. Every design rule is tied to the paper that motivates it, with tiers and links, and every component is mapped to a Stories primitive with the platform gotchas we hit along the way.
Would love to hear if others have run into the same wall, and how you handled it.
deterministic-sandwich-cross-run-learning-tines.pdf (290.7 KB)