Training Agent Writes MD Spec, Deployment Agent Codes Inference: A New Workflow Under Fire
ChenTessler · x · 2026-10-08
The substance is in the quoted proposal: your training-side agent writes an MD file describing in detail what the policy observes and outputs; your deployment-side agent reads it and codes the inference code.
ChenTessler pushes back that this sounds error-prone — everything that can be compiled should be compiled, with a single source of truth, rather than having agents write specs for other agents to reimplement in C++ with testing infrastructure to verify equivalence. A methodological debate over deterministic compilation vs agent-generated pipelines in robot policy deployment.
Related event: NVIDIA ProtoMotions Workflow Sparks Debate(2 posts)→
More from coding & agent
- Agents now open the exact app screen where their change lands, cutting eval legwork — lucasmeijer · 2026-10-09
- Dev shows remote workspaces running any agent on any model, incl. vanilla Claude Code — lucasmeijer · 2026-10-09
- Open-source lithos-metal hits 200+ tokens/s/user on Qwen3.8-27B with one M5 Max — JiaZhihao · 2026-10-09
- Remote dev setup: each workspace gets its own agent, browser, and terminals — lucasmeijer · 2026-10-09
- Claude Code 2.1.294 fixes instruction-style hooks and premature agent stops — ClaudeCodeLog · 2026-10-09
- Claude Code 2.1.294 fixes instruction-based hooks that failed to block commands and premature agent stops — ClaudeCodeLog · 2026-10-09