How to Build a Model Router in the Harness
Open SWE's harness-level model router cut median cost per coding task 64% with no measurable quality drop; the post shows how to build one.
AI Daily Recap
28 articles · 4 categories
The finishable daily brief
What happened in AI — Oct 2, 2026
Friday, Oct 2, 2026
28 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- Open SWE's model router cut median cost per coding task by 64% with no measurable quality drop.
- Docker brings the Sandbox Kit Spec to the CNCF: agent permissions packaged as OCI images.
- OpenAI DevDay: GPT-6.1 Sol, computer use in the Agents API, cloud Codex environments.
- Uber Eats cut search latency 50%; AWS shows RL fine-tuning for small search agents.
- A Chinese open-weight model is under investigation over bioweapon-instruction claims.
Agent cost and safety are being decided in the harness. Open SWE's router cut median coding-task cost 64%, and Docker is taking sandbox permissions to the CNCF as OCI images.
OpenAI's DevDay adds GPT-6.1 Sol and computer use to the Agents API, while local hardware and open-weight safety probes show where self-hosted agents stand.
Agent Harnesses, Routing, and Managed Runtimes 5 items
Harness-level design is where agent cost and safety are now decided: model routing, sandbox permissions, and managed microVMs.
Docker Sandbox Kit Spec: Packaging AI Agent Permissions as OCI Images
Docker is taking its Sandbox Kit Spec to the CNCF, packaging what an agent may access as an OCI image that travels with the agent.
DigitalOcean Managed Agents Brings Managed Cloud Infrastructure to AI Agents
DigitalOcean Managed Agents entered public preview, running agents in isolated microVMs on managed infrastructure.
[AINews] Pi 1.0, Pi Durable, and AIE NYC
Pi 1.0 ships as a stable, TypeScript minimalist harness, alongside Pi Durable for long-running work.
Show HN: What's Agent Doing – a Claude Code UI mod that explains each step
A Claude Code UI mod that explains each step, aimed at the opaque shell commands models now run in deep-work mode.
Frontier Model Releases and Agent Tooling 4 items
OpenAI's DevDay and GPT-6 guidance push reasoning-effort tuning, computer use, and hosted Codex into the production agent path.
OpenAI DevDay 2026 Recap for Developers
DevDay 2026 brought GPT-6.1 Sol, computer use in the Agents API, and cloud-based Codex environments.
A model guide for the GPT-6 family
OpenAI's GPT-6 guide covers choosing a model, tuning reasoning effort, and coordinating tools and skills for production workflows.
Chatham scales its capital markets expertise with OpenAI
Chatham Financial used Codex and GPT-5.6 to cut trade validation from 30 minutes to under 4.
How to choose your first Genie Agents for maximum impact
Databricks says more than 1 million Genie Agents were created in 2026 and offers guidance on picking first use cases.
Training, Latency, and Production Case Studies 3 items
Teams report measured gains from fine-tuning small agents and rebuilding pipelines rather than adopting bigger models.
Fine-tune a search agent with multi-turn RL on Amazon SageMaker AI
AWS shows multi-turn RL on SageMaker AI teaching a small search agent your tools, targeting frontier-model reliability at lower latency and cost.
Uber Eats Rebuilds Search Pipeline to Cut End-to-End Latency by 50%
Uber Eats rebuilt its search pipeline and reports a 50% cut in end-to-end latency, with new Above-the-Fold measurement.
Claude Frontier Academy: $100M to train 10,000 engineers
Anthropic committed $100M to train 10,000 Frontier Deployed Engineers by the end of 2027.
Local Inference and Open-Weight Safety Scrutiny 5 items
Local hardware makes open models practical for coding agents, while Chinese open-weight models draw safety and censorship scrutiny.
NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI
NVIDIA's DGX Spark 64GB targets running increasingly capable open models locally for agent development.
HP ZGX Nano G1n Review: GB10 Mini PC Runs Local Qwen Coding Agent
HotHardware's review has a GB10 mini PC running a local Qwen coding agent.
Chinese AI model investigated after researcher says it provided instructions for bioweapons, assassinations
A Chinese AI model is under investigation after a researcher said it gave instructions for bioweapons and assassinations.
Some topics are off limits inside popular Chinese-made free AI, researchers find
Researchers found some topics are off limits inside popular free Chinese-made AI models.
DeepSeek, Huawei Join Forces to Challenge Nvidia’s CUDA Dominance
DeepSeek and Huawei are reported to be teaming up to challenge Nvidia's CUDA dominance.
You are caught up for this edition