Open SWE’s model router cuts median task cost 64%

#1 𝕏

Harrison Chase recapped that Open SWE moved model choice into the harness, routing each task to the cheapest model that met tested quality requirements. He said median cost per task dropped 64% and agreed that most orchestration steps do not need a frontier model.

#2 𝕏

Garry Tan shared his gbrain-evals repository, which includes 40+ categories that can each be auto-researched and auto-improved. He said clear evaluations for retrieval, latency, and cost per compute are useful for retrieval and latency data tasks like those performed by GBrain.

#3 in

Every Inc. shared how Dan Shipper built a copyediting skill from 30,000 historical edits and installed it in a company agent operating in Slack. Every Agent, Every Inc.’s AI coworker in Slack, lets teams request tasks in channels, correct outputs in threads, and save workflows as reusable skills.

#4 𝕏

Boris Cherny said Claude has handled an unspecified task better than he could since around Opus 4.8, and that 5.5 models are now very good at it. His workflow is to provide a goal, a way to measure results, and guidance on effort and duration—potentially a few weeks. He often also requests a daily progress chart or a live artifact with the result.

#5 𝕏

Sebastian Raschka shared round 2 of “Reasoning From Scratch: Reinforcement Learning with Verifiable Rewards (RLVR),” covering clipped policy ratios, a KL loss term, format rewards, and other GRPO tips and tricks. The original post includes detailed chapter timestamps on GRPO training, metrics, entropy, policy ratios, KL divergence, and format rewards.

Also covered by: @Sebastian Raschka

#6 𝕏

Garry Tan commented that Capy.ai works best with experimental modes, including Captain, turned on. He uses one thread to coordinate the merge PR queue, ensure CI works, and allow only green runs with full CI into master.

#7 𝕏

Guillermo Rauch shared that 58.18% of traffic across the Vercel network was bot-originated over the last 30 days, versus 32% in January 2024; more than 60% of Vercel deployments are now agentic, up from approximately 3% in January 2026. Agents also account for up to 83% of pageviews to Vercel’s documentation sites, with that share increasing as Vercel optimizes content for agents.

#8 𝕏

Harrison Chase shared @ch3nweiii’s framing that many agent steps are yes/no calls rather than generation: Jev handles them as typed questions with calibrated probabilities, leaving harder work to the frontier model. LangChain’s “Building a Harness with Jev” explains where those calls fit within a harness.

#9 𝕏

Boris Cherny commented that API usage typically involves Agent SDK-shaped workloads, where a 1h TTL would raise average costs, so the default remains 5m. Interactive usage defaults to 1h because it is cheaper for users on average, though the configuration can be changed.

#10 in

Udi Menkes recaps Noah Shinn, founder of Instinct, demonstrating Instinct’s hospitality experience on October 9 (year unspecified): it generates three different hotel-booking interfaces based on location, price, or family needs, then books the selected room. For AI product builders, Menkes argues for defining components, signals, assembly rules, and constraints so agents can tailor interfaces to each user’s intent.

#11 𝕏

Lenny Rachitsky shared an emerging “row, then steer” playbook for PMs at large companies exploring zero-to-one ideas: a PM and designer vibe-code a working alpha, bring in engineers once it shows pull, then step back from coding to align, prioritize, and unblock. The post draws on a story from Atlassian CPO @TYehoshua.

#12 𝕏

Madhu Guru said agent capabilities will soon evolve at an “RSI pace,” requiring agentic security—including identity, permissions, and observability—to advance even faster. Guru characterized the agents space as progressing from capability to trust to autonomy, with the focus shifting toward whether systems take appropriate actions with appropriate permissions and can be observed, audited, and held accountable.

#13 𝕏

DeepLearning.AI shared Andrew Ng’s argument in Andrew’s Letter in The Batch that agents use computers to search, store, analyze, and transact at unmatched scale, requiring much more infrastructure and faster tools—not just faster models.

#14 𝕏

clem 🤗 commented that total usage of the hub is growing, with daily active users considered more trustworthy than raw download numbers. They questioned whether @natolambert’s graph tracks only 2,000 models and said total model downloads have been unreliable and are not closely followed.