Tag
Briefings and ideas about Agent harnesses from AI podcast discussions and the research they cite, with links to the original sources.
Modal CEO Erik Bernhardsson says the team once discussed shutting down sandboxes, its service for running AI-written code. They now bring in about a third of revenue. He expects agents to move toward brief, snapshotted environments.
Tenex, which builds AI systems for large companies, says it stores plans, agent decision logs and project context beside its code so agents can read them. In the episode, Tenex said Opus 5.5 is the main coding model for most of its team, and that pairing it with Codex as a reviewer still gives the best results.
Runtime Verification's leaders say an LLM's Rust port of kernel code, checked by differential fuzzing and Lean proofs, surfaced Linux bugs. The company's write-ups show partial proofs that are only as good as their specifications.
Dave Blundin said today's AI forgets mistreatment like a goldfish once its context is flushed. On Moonshots, Alex Wissner-Gross countered that tools like Claude Code already save notes to files, so "I wouldn't count on it."
Dyna Robotics' Jason Ma says hour-long robot jobs such as laundry need a planning layer around the low-level action model. He names fragile robot hands, slow hardware testing and vision-only data as limits on more general work.
Goodfire's Eric Bigelow argues in-context learning covers nearly everything a model does to adapt as it reads. He expects context, memory and harnesses, not per-user fine-tuning, to drive personalization, but hasn't researched continual learning.
A preprint credits an unreleased Claude model with the seed of a 3SUM algorithm just below quadratic time. On Moonshots, the result was cited in an argument that labs' internal models are pulling away from open ones.
OutSystems CEO Woodson Martin says the company's AI token spending peaked in June and July, then fell below forecast after it built its own harness and model router. He argues most enterprise work doesn't need frontier models.
Yash Patil, whose company Applied Compute sells post-training, says teams should first improve prompts, context and tools, then retrain weights for their biggest workloads, tasks that change over time and judgment-heavy tasks.
Yash Patil, CEO of Applied Compute, argues that the biggest AI workloads are the most valuable both to custom-train and to serve, which he says gives a training provider an edge in inference. He calls serving the easier job.
Alex Zhang, an MIT PhD student, described Prime Agent, built with Prime Intellect: the agent acts only through Python and keeps its history on disk. He also cited outside RLM work, including Harvey's legal agents and Laude Institute's Headlong.
MIT's Alex Zhang says Claude Code, Codex and Pi share nearly all their design. He argues code-based recursive harnesses could help models generalize, citing his own early experiments on one open model as support.