If this agent really learned from its own failures, “just add more context” is officially dead.
We thought our agent was nondeterministic. It wasn’t. It was consistently wrong in ways we couldn’t see—until we added Hindsight.
We built a tool-using agent and wired in Hindsight to record + replay every run.
Here’s what actually changed:
• Before: same input → different tool choices → random failures
• After: same input → same decisions → stable outputs
Not because the model changed. Because the state stopped drifting.
• We stopped treating memory as “more tokens”
Instead, we stored full execution traces: inputs, tool calls, outputs.
• We normalized tool responses
This alone removed most “randomness” (LLMs hate inconsistent schemas).
• We replayed failed runs
Hindsight showed exactly where decisions diverged—step by step.
• We fed those failures back in
The agent learned patterns like:
“Don’t retry empty results”
“Prefer lookup over search when key exists”
• Behavior actually changed over time
It stopped looping. Stopped picking the wrong tool. Became predictable.
This wasn’t RAG.
This wasn’t bigger context.
This was experience → feedback → better decisions.
If you’re building agents, the takeaway is simple:
They don’t need more memory. They need usable experience.
Save this if you’re about to bolt memory onto your agent stack.
What’s the most surprising thing your agent has “learned” from its own failures
Top comments (0)