/Stream
All 1279 resources sorted by date, newest first
Are harnesses worth the cost?
Ten runs on a Java API service show a behavioral harness lifting a weaker model from three to ten successes at lower cost than a stronger unharnessed model, with caveats
Gemini 4 Argon
Google DeepMind's new flagship model for complex software engineering, research and robotics, with a one-million-token output window
How to effectively debug when you no longer understand the code
Three production investigations, a latency regression, a memory leak and a kernel-level hang, show how to steer coding agents through unfamiliar systems by separating facts from hypotheses and comparing failures with a control
Muse Spark Open
Meta's open-weight multimodal MoE model with 816B total and 42B active parameters and a one-million-token context, runnable on two M5 Ultra Mac Studios
Speculative decoding, the art of guessing well
Deep walkthrough of drafting several tokens cheaply and verifying them in one target-model pass, with acceptance and speedup formulas, EAGLE-3, multi-token prediction and diffusion drafters
Why WebMCP matters
Sites expose JavaScript tools to agents instead of being driven through screenshots and DOM inspection, with a Shopify checkout experiment running 2.7 times faster at 58% lower cost
AI Policy
Community collection of ready-made contribution policies for open source projects, from welcoming AI contributions to human-only
Bifrost
Open-source AI gateway with an OpenAI-compatible API, provider failover, load balancing, semantic caching, virtual keys, budgets and MCP support
OpenCode 2.0
Big update to the open coding agent with background subagents, a code mode that lets models script MCP tool calls, and JS and Effect clients to drive it
REA
Gives coding agents tools to reverse engineer native binaries, JavaScript and Electron applications and browser behavior
SPACE
Perplexity open-sources the sandbox platform behind Computer as a foundation for safer agent systems
Agents don't need memory. They need documentation.
Vector-search memory surfaces similar snippets without proving they are current or relevant, while a structured Markdown workspace of specs, decisions and instructions gives agents auditable context
Claude Haiku 5.5
Fast, low-cost model for summarization, classification, browser use and subagent tasks, with average running costs about 75% lower than Haiku 4.5 and adjustable effort
Don't merge what you didn't run
A test gate for agent-written code that reinstalls from scratch, runs the full suite three times, rejects removed or skipped tests and restores the base branch's tests against the new code
Ephemeral testing
Test a software component by having an agent build disposable applications on top of it, exposing hidden state, unstable invariants, surprising defaults and weak documentation
GPT-6 and Intelligent UI for everyone
GPT-6 rolls out in ChatGPT with Intelligent UI, which combines text, graphics, forms, charts and interactive tools in a response and can start answering while still reasoning or searching
How ChatGPT Intelligent UI works
Reverse-engineering tour from a model-written DIL format through server-side compilation, sandboxed client execution and native rendering, including how streaming recompiles partial output and preserves keyed state
Introducing Beam, Reflection's 501B open-weight model
Sparse Mixture-of-Experts model with 501 billion total and 23 billion active parameters aimed at coding, reasoning and agentic workloads, with Apache 2.0 weights planned for later this month
Introducing Mistral Large 4
One-trillion-parameter multimodal model in public preview with 49 billion active parameters, 160+ languages and open weights planned for later this month
Introducing the Personal Agent Protocol
A new protocol for personal agents that is built on top of OAuth