Build an agent that can keep improving
Engineering field guide · Agent systems · Both days
What should persist between agent runs?
Connect traces, memory, triggers, and recovery. Then use that feedback to improve the model as well as its runtime context.
All topic guides · Practical takeaways
Where the approaches meet
Section titled “Where the approaches meet”The talks connect agent quality to a feedback loop. Traces, saved state, and human decisions give the next run information the previous run lacked.
How the approaches differ
Section titled “How the approaches differ”| Session | Approach |
|---|---|
| Arize AI · Day 1 | Signal reads traces and proposes fixes. Issues and learned skills persist while temporary access is removed after each run. |
| Google DeepMind · Day 2 | Focuses on managed execution: triggers, protected hooks, and credentials kept behind a proxy. |
| Prime Intellect · Day 2 | Extends the loop toward model training. Traces and approvals can become evaluations and training labels. |
Arize AI · Day 1 · Google DeepMind · Day 2 · Prime Intellect · Day 2
Where the evidence stops
Section titled “Where the evidence stops”Persistent memory and model training solve different problems. Training adds data, evaluation, and operating costs. The proposed overnight learning loop is not evidence that every organization benefits from daily training.
Try this on your workload
Section titled “Try this on your workload”- Choose a recurring task with a clear trigger and save its run state outside the sandbox.
- Interrupt a run and test whether it resumes without repeating completed actions.
- Review ten runs and turn one recurring failure into a regression case before changing the agent.
Measure the result
Section titled “Measure the result”Track duplicate actions, recovery success, repeated mistakes, and time spent reviewing fixes. Use the baseline to decide whether a memory change helps.
Read the source sessions
Section titled “Read the source sessions”- Arize AI · Day 1. Our session notes · Official reading edition · Source video
- Google DeepMind · Day 2. Our session notes · Source video
- Prime Intellect · Day 2. Our session notes · Source video
The comparison and experiment are our editorial synthesis. They are not a reported benchmark or the speakers’ prescribed procedure.