Akilesh K R

@aki07

Retry or Retrain? Building a Verifiable RL Environment for Coding Agents

Submitted Oct 31, 2026

Abstract

When a coding agent fails, giving it feedback and another attempt is often the simplest next step. Training the model requires additional effort and compute, but tools such as OpenEnv and Hugging Face TRL make it possible to explore that choice. The harder question is whether training improves outcomes enough to justify the cost. Before answering that, we need to trust the signal we train against. Passing tests does not always mean fixing the underlying problem. An agent might hardcode expected outputs or satisfy visible examples while failing on other inputs. How do we check that our reward measures the behaviour we actually want?

This talk walks through building a resettable reinforcement learning environment from real bug fixes in open-source Python repositories. A small coding agent can search and read code, edit source files, run visible tests, and submit a patch. Each attempt starts in a fresh, isolated environment. Protected bug-specific tests and regression checks determine whether the submitted patch succeeds.

Attendees will leave with a way to decide whether post-training is worth trying on their own agent. The talk shows how the agent, environment, verifier, and training loop fit together, and how to check each piece before spending compute. It comes from an independent study, so it presents a method to adapt rather than a proven result.

Key takeaways

  • How to turn a bounded, tool-using workflow into a resettable RL environment.
  • How to audit rewards using deliberately injected shortcuts.
  • Why injected exploits and agent-discovered shortcuts should be reported separately.
  • How to compare training against feedback and retries under consistent budgets.
  • What evidence to gather before applying the approach to a production agent.

Target audience

Engineers building tool-using agents, ML practitioners interested in post-training and evaluation, and technical leads assessing whether training is worth the investment. No prior RL experience is required.

Format

30-minute talk. Slides will use code excerpts, execution traces, and screenshots to walk through an agent episode and a reward-audit example.

Speaker bio

I am Akilesh KR, a Solution Consultant at Sahaj Software Solutions. I enjoy exploring machine learning, building things to understand how they work, and sharing what I learn. I contribute to open source and write about AI and technology on multiple forums.

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

Jumpstart better data engineering and AI futures