Sign InOpen Brain
AI EngineerVideoSource Linked

Learning on the Job: The Future of Post-Training — Raymond Feng, Applied Compute

Training on a production harness could adapt an agent to real workflows, but non-replayable interactions break common RL assumptions and make feedback harder to turn into updates.

AI Engineer · Aug 1, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Replayable synthetic environments support **GRPO** by running several trajectories from the same initial state, grading them, and adjusting model weights toward the better paths.

Practical Implication

For a custom agent, preserve traces and feedback at the model boundary. A **bring-your-own-harness** setup can train on the actual workflow instead of a brittle simulation, even when the training team cannot inspect its source.

Agent-Ready Context
Replayable synthetic environments support **GRPO** by running several trajectories from the same initial state, grading them, and adjusting model weights toward the better paths.

For a custom agent, preserve traces and feedback at the model boundary. A **bring-your-own-harness** setup can train on the actual workflow instead of a brittle simulation, even when the training team cannot inspect its source.

Production interactions may be non-replayable and off-policy, so conventional comparative rollouts no longer fit. **Self-distillation**, automated trace curation, and qualitative-feedback ingestion remain open research directions rather than general solutions.
Connected Context · Feed7 Judgment

This turns production traces and harness feedback into potential post-training material, extending self-improvement from issue discovery toward weight updates on the real workflow. It also sharply limits that promise: GRPO depends on replayable comparable trajectories, while live interactions are often off-policy and non-replayable, leaving trace selection, qualitative feedback, and self-distillation unresolved.

From Signal to PR: Anatomy of a Self-Improving Agent — Jason Lopatecki, ArizeArize shows the nearer-term use of production traces for investigated issues and draft fixes; this talk extends the same evidence stream toward model training while preserving the need for human review.Data and Environment Curation for Post-Training LLMs — Mahesh Sathiamoorthy, Bespoke LabsIts emphasis on rollout quality, filtering, task mix, and environment design supplies prerequisites for turning captured harness traces into useful post-training data rather than assuming RL will help.Reinforcement Learning without Verifiable Rewards — Will Brown, Prime IntellectBoth use production traces and simulators to bridge deployment and training, but the candidate adds judge-derived rewards and reward-hacking risks where deterministic grading is unavailable.Why Off-the-Shelf AI Doesn't Understand Money — Udi Menkes, IntuitIntuit’s verified state-action-outcome histories illustrate the kind of grounded trajectory record this approach needs, while also showing that useful learning depends on valid outcomes and well-matched evidence.
Context Map
agent#harness-engineering#agent-reliability#tool-use
Uncertainty
Production interactions may be non-replayable and off-policy, so conventional comparative rollouts no longer fit. **Self-distillation**, automated trace curation, and qualitative-feedback ingestion remain open research directions rather than general solutions.