Personal AI News Digest
16 topics from 3 sources · Archive
1,389 words · ~7 min read
Loop engineering—designing AI agent workflows that run on schedules or event triggers—has emerged as the dominant paradigm for production agent systems. The term gained prominence in mid-2025 when engineers at Anthropic and OpenAI (Boris Cherny, Peter Steinberger, Addy Osmani) shifted focus from prompt engineering to loop design. Geoffrey Huntley's "Ralph Wiggum" loop pattern breaks large tasks into smaller agent runs to work around context window limits (~200k tokens). By May 2025, major coding harnesses—Codex, Hermes, and Claude Code—shipped /goal and /loop commands making loop engineering accessible via single commands, similar to JavaScript's setTimeout(). Common use cases include development workflows (opening PRs for app issues, fixing flaky tests, triaging issues, reviewing design plans), nightly/daily work (product improvements, end-to-end test runs), and complex migrations. The outer loop includes human oversight, feedback signals, and evaluations; agents run the inner execution loop while engineers set direction and make decisions in the outer loop. Dex Horthy from HumanLayer cautioned that "the hype is outrunning the discipline," noting that Kubernetes loops are deterministic, unlike agent loops.
Sources Latent.Space, The Pragmatic Engineer
Links the first morning of keynotes at AIEWF, an interview with Latent Space, an on-stage debate was held, What is “loop engineering?”, Gergely Orosz, said, Peter Steinberger, an article, Ralph Wiggum as a software engineer, Ship working code while you sleep, the documentation, shipped, shipped, /loop command, /goal plugin, n8n, cron jobs, Oded Messer, Ivan Pantić, Paul D’Ambra, Ivan Abad:, Artem Nikitin, Jack D,, Lawrence Jones, Rafel Mendiola, Aaron Stannard
The field has matured from 2023's focus on autonomous agents (AutoGPT, BabyAGI) to engineering reliable systems that manage workflows, context, permissions, evaluation, and continuous improvement. Lilian Weng's 2023 essay on LLM-powered autonomous agents contrasted with her 2026 essay on harness engineering, reflecting the shift. Complete agent autonomy has proven unreliable and undesirable at scale; agents now augment engineers rather than replace them. Anthropic's Thariq Shihipar noted that models are "grown, not designed" with "capability overhead" where "Claude gets smarter in a spiky way," making evaluation and monitoring systems essential. As code generation costs trend toward zero, quality and observability become first-class differentiators. Stale agent instructions can cause multi-hour stalls in long-running tasks. LangChain added tracing for Codex, Cursor, Copilot, Pi, and OpenCode in LangSmith, exposing tool calls, subagents, and token usage. Hermes added parallelization of tool calls and banked resets.
Sources Latent.Space, AINews
Links LLM Powered Autonomous Agents, Harness Engineering for Self-Improvement, In a separate keynote, Thariq Shihipar talked about how their latest model, Claude Fable, Richard MacManus, 5 Trends That Defined AI Engineering at World’s Fair 2026, Addy Osmani’s excellent keynote, @swyx warned, LangChain added tracing for Codex, Cursor, Copilot, Pi, and OpenCode in LangSmith, @Teknium shipped Hermes updates, banked resets directly in Hermes Agent, @andykonwinski
Tools like Claude Code, Codex, Gemini CLI, Cursor, and Warp now handle broader objectives: exploring codebases, modifying multiple files, running tests, debugging, and iterating before presenting results, replacing 2023's GitHub Copilot-style line completion. Codex usage jumped from 5M to 6M active users in approximately 24.5 hours following Tibo's announcement, potentially overtaking Claude Code (which had 2M users in February) as the leading coding assistant. Vercel released eve, an agent framework comparable to Next.js. Andrew Qu (Vercel Chief of Software) noted agents are "not as predictable as web applications" with "much more dynamic" interaction and outputs, and that a year ago, sandboxes and secure code execution weren't recognized as critical; production experience continues to reveal infrastructure gaps. Charlie Holtz (Conductor) cautioned that engineers should remain in control, preferring orchestration metaphors over factory automation.
Sources AINews, Latent.Space
Links Yesterday’s headline story, told Latent Space at AIEWF
Anthropic popularized "agent skills" (introduced October 2025) as a way to encode workflows, quality gates, and best practices. Addy Osmani defined skills as portable, on-demand knowledge. Roland Gavrilescu (Introspection) declared the field has shifted "from agent tools to agent skills." Philipp Schmid (Google DeepMind) demonstrated using declarative Markdown skills to enable "agents without code," reducing orchestration code previously written in Python. Paul Bakaus created Impeccable, an open source design skills system for coding agents. Matt Pocock warned of "skills hell" (analogous to frameworks hell) and provided a checklist: write fewer, smaller skills with better structure. Y Combinator president Garry Tan urged startups to encode business functions as skills and hire engineers to maintain them. Tyler Brown cautioned that "autonomy without structure creates as much slop as leverage."
Sources Latent.Space
Links last October, Addy Osmani’s definition, Impeccable, an interview with Latent Space, noted on X
A new role called "Forward Deployed Engineer" (FDE) is emerging where engineers work directly with organizations to implement AI capabilities. Natalie Meurer (Sierra) noted that enterprises need to "manage all the integrations and all the teams that contribute to the agent." Cursor's Pauline Brunet described implementing "software factories" with "long-running agents helping people throughout that entire process." Warp CEO Zach Lloyd explained that organizations must choose which parts of the software lifecycle to automate and where humans should intervene. Prukalpa Sankar (Atlan) emphasized "context engineering"—how context flows from business systems into agents via MCP and APIs. Enterprise adoption remains concentrated among early adopters.
Sources Latent.Space
Links Forward Deployed Engineers and the future of software engineering, Zach Lloyd explained, a tweet, told Latent Space at AIEWF
Agent Arena demonstrated system cost reductions of 89% while matching best static config accuracy, arguing full system config outweighs LLM routing alone. Google DeepMind work on model routing argued routers should be judged by behavioral differentiation among experts and stability under paraphrase, not just accuracy/cost. Hamel Husain's automated evals post noted these systems spot issues humans miss but lack domain taste and feedback loops to replace experts. Benchmarks are expanding beyond one-shot tasks: mini-swe-agent marked one year while powering multiple software benchmarks; SlopCodeBench measures how agents erode codebases over sequential tasks rather than solving isolated issues.
Sources AINews
Links Agent Arena, Google DeepMind work on model routing, @HamelHusain’s automated evals post, mini-swe-agent, SlopCodeBench
Perplexity open-sourced WANDR, a benchmark built from 500 de-identified production research tasks with 170,495 source-backed records across multiple difficulty tiers. Rather than grading against static gold sets, WANDR re-fetches cited pages and checks claims against underlying evidence, better matching dynamic web research. Framed as the internal benchmark behind Perplexity Computer's research harness and as an RL environment synthesized from production traces.
Sources AINews
Links @perplexity_ai, @AravSrinivas, @denisyarats
OmniAgent, built on Qwen2.5-Omni-7B, uses an Observation–Thought–Action loop to request only needed frames and audio rather than ingesting all frames. On LVBench, OmniAgent-7B scored 50.5, beating Qwen2.5-VL-72B at 47.3 while consuming only ~203 frames vs 768. Training used 58K agentic trajectories and entropy-weighted RL via TAURA; passive SFT hurt performance. The research aligns with the view that motion is a fundamentally novel data type requiring dedicated collection and model treatment.
Sources AINews
Links @ZhihuFrontier, Andrew Carr’s note
OpenMOSS released MOSS-VL-Realtime, an 11B vision-language family under Apache 2.0 with 256K context, designed for continuous video streams. The model can keep watching while generating, revise or interrupt answers as scenes change, and remain silent when evidence is insufficient. Architecture features include cross-attention, XRoPE for unified temporal-spatial positioning, and unified templates across offline/streaming/realtime settings.
Sources AINews
Reka AI outlined the data stack behind omni world models, stressing petabytes of video, 6 pipeline stages, and doubled payoff from data-quality improvements when models both generate and understand video. LingBot-World 2.0 was highlighted as one of the first open releases claiming hour-scale, 720p/60fps interactive generation, though without long-term memory. PixVerse Game is pursuing real-time interactive video response rather than canned clips.
Sources AINews
Links @RekaAILabs outlined, @omarsar0 summarized LingBot-World 2.0, PixVerse Game
PrismML released Bonsai 27B, based on Qwen 3.6 27B, in two variants: Ternary Bonsai 27B at 5.9 GB / 1.71 effective bits and 1-bit Bonsai 27B at 3.9 GB / 1.125 effective bits, both under Apache 2.0. The model preserves multimodal, tool-using, and long-context agentic workflows locally; demos show Hermes running on RTX 5090 and phone deployment.
Sources AINews
Links PrismML released Bonsai 27B, a demo shows Hermes running it on an RTX 5090, Locally AI highlighted phone deployment
Tencent released 1-bit and 4-bit quantized versions of Hunyuan Hy3, a 295B flagship-scale model that can be served on a single GPU via llama.cpp with MTP enabled.
Sources AINews
Daniel Hanchen announced NVFP4 dynamic quantization support across the Gemma-4 family and additional large models including Qwen3.5-122B-A10B and GLM-4.7-Flash. MiaAI_lab's DGX Spark thread sketched practical multi-node local deployments: 1M-context DeepSeek v4 Flash and MiMo-V2.5 on 2× DGX Sparks, and GLM 5.2 NVFP4 across four nodes.
Sources AINews
Links @danielhanchen announced NVFP4 dynamic quants, @MiaAI_lab’s DGX Spark thread
Sakana AI published Smart Cellular Bricks in Nature Communications—identical cubes running small neural networks, communicating only with physical neighbors, yet able to infer global shape and detect damage without centralized control. Cells detect missing neighbors across six spatial directions with 95% accuracy and can regrow target structures; the method scaled to 18,000+ cubes in simulation.
Sources AINews
Links Sakana introduced “Smart Cellular Bricks”, detail thread
An autonomous micro-drone achieved an air-to-air kill of a flying moth, framed as a step toward mosquito eradication.
Sources AINews
Links @alextoussss posted
Airtap turns SMS into a headless agentic execution layer for mobile apps, using text as the control plane and intervening only for authentication.
Sources AINews