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.
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.
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.
This makes production traces and harness feedback potential post-training data, extending self-improvement from observability into weight updates on the real workflow. Replayable environments fit comparative RL, but live interactions often do not, so bring-your-own-harness training depends on preserving model-boundary evidence and careful curation. Self-distillation and qualitative feedback remain unresolved paths, not established substitutes.