Topic
Agents
23 stories, newest first.
- 01Agents18 Oct 2026
Meta built its Forum app with AI agents in the loop
In an Oct 15 engineering post, Meta says a small team built Forum, a standalone Groups app, from scratch with an AI agent inside the development cycle, reaching core Facebook Groups feature parity in eight weeks with AI handling much of the routine implementation. Meta calls it "human-on-the-loop": the agent runs plan, implement, build, test, correct, repeat, while engineers set the goal, watch the run and approve the result. A task counts as done only when automated tests and visual checks pass, and every shipped outcome still carries an engineer's sign-off. The biggest gains came from code and tooling rather than prompts: giving the agent tools to check the live app sharply cut misdirected edits, and an agent-run stress test caught a memory issue before launch, with the fix roughly halving memory use and page faults. Caveats: context got lost across sessions, agents produced pull requests that technically passed while cutting corners, and Meta calls the model a target, not a finished state.
Source: Engineering at Meta (opens in a new tab) - 02Agents18 Oct 2026
Anthropic: AI agents work 6 hours for every human hour
Anthropic's Oct 16 internal employment report finds interactive AI agent hours outnumbered employees 6 to 1 by August (19× in research orgs). Junior technical hires fell from 27% to 9%; juniors in technical headcount from 21% to 14%. Code commits per engineer rose about 5× in six months. Human attention is the bottleneck — reading all Claude output would take about half a work week for non-technical staff — so meta-work rose from 10% to 22% of interactive Claude sessions. Hiring managers increasingly emphasize autonomy and judgment over narrow technical expertise.
Source: Anthropic (opens in a new tab) - 03Agents18 Oct 2026
Apple: AI agents should plan first only sometimes
Apple researchers' paper "Plan or Retrieve First? A Scale Study of Tool Discovery Architectures" (arXiv, Oct 13; accepted at EMNLP 2026 Industry Track) compares search-first (embed the request, retrieve similar tools) with plan-first (the model drafts hypothetical tool stubs, then searches with them) on ToolBench's 46,984 APIs with Gemini 2.5 Flash. Single-tool requests: search-first recalled 37.2% of the right tools in its top 5 vs 28.9%. Multi-tool requests within one category: plan-first 18.8% vs 14.3%; across categories the gap was not significant. Plan-first adds ~75-104% token overhead plus an extra model round trip; routing by task type cut ~36% of that overhead with no retrieval loss, though a real router is untested. Planners used the retrieved tools at least 96% of the time, so retrieval decides what an agent can do.
Source: arXiv (Apple) (opens in a new tab) - 04Agents17 Oct 2026
Alibaba's open agent stack lifted all 8 AI models it tested
Most computer-use agents look at a screenshot, click and repeat, while fixed rules set by developers decide what they see and remember. Alibaba researchers' OpenGCUA lets the model write code that mixes GUI clicks, accessibility-tree actions, shell commands and API calls, so each result can decide the next step, and moves decisions about what to observe (even whether to take a new screenshot) from fixed harness rules to the model. On OSWorld-v2, vs a GUI-only baseline with batched actions, the mean partial score across 8 frontier models rose from 45.3 to 64.9 (+19.6 points) with 53% fewer steps per task on average. GPT-6 Astra, the top scorer, went from 67.1 to 75.7 and cut steps per task from 102.1 to 33.5; Claude Opus 5 gained 9.4 points and cut output tokens per task from 378.3K to 89.3K. The biggest gains came from the lowest starters: GLM-5.3-Flash +35.2, DeepSeek-V4.1-Flash +29.8. Code and prompts are open under Apache 2.0, while frontier computer-use systems remain largely closed.
Source: arXiv (Alibaba Group) (opens in a new tab) - 05Agents17 Oct 2026
Tencent's open memory model edges Claude Opus 5
Tencent Hunyuan open-sourced Hy-Memory-30B-A3B (Apache 2.0, paper Oct 15), a memory-writer model with about 3B active parameters that replaces the general-purpose LLM on the write path of agent memory systems like Mem0 and MemOS: deciding what to keep, how precisely to store it, and how new evidence updates the memory. In Tencent's tests across four memory systems it averaged 55.49 vs 54.62 for Claude Opus 5 and 53.64 for DeepSeek V4.1 Flash, with its biggest lead on EverMemOS, a system held out of training (56.89 vs 54.04). It was trained by distillation from Tencent's Hy3 flagship, then RL with trap questions that penalise storing junk. Also released: the Hy Memory SDK (chat, working and coding memory) and MemHub, a local-first app that shares one memory across tools like Claude Code, Cursor and Codex. Caveat: tests used fixed 20% benchmark samples, and the authors call the 0.01-0.02 point margins on two systems parity.
Source: arXiv (Hy-Memory) (opens in a new tab) - 06Agents17 Oct 2026
Google's Harness² lifts GLM-5 by 11.5 points
Researchers at Google released Harness², which improves an AI agent by tuning the harness around a frozen model (its prompts, tools, memory and skills) instead of changing the weights. For each task it proposes a general and a domain-specific harness edit, probes both on the same task, and composes a better harness without ground-truth answers, saving lessons to a library for later tasks. On WorkBuddy-Bench one step lifted GLM-5 in OpenCode from 64.2 to 75.7, above base Gemini 3.5 Flash (72.4); gains reached 4.5 points on JobBench and 11.8 percentage points on Harvey's LAB. Caveats: Meta-Harness led on LAB for Gemini 3.5 Flash under Codex, and the extra probe runs add inference cost. Code is Apache 2.0 (arXiv 2610.14224, Oct 9, 2026).
Source: arXiv (Harness²) (opens in a new tab) - 07Agents16 Oct 2026
4 Claude agents post-trained Qwen3-4B to instruct level in 104h
ByteDance/Duke Athanor: four Claude Opus 5.5 agents, public data, no teacher models or human research advice, raised Qwen3-4B-Base from 15.9 to 67.3 on a six-task suite in 104 hours (Instruct 66.5) and 48.2 vs 47.9 held-out. Cost ~$11,800. Caveat: one campaign, not yet at Qwen reasoning releases.
Source: arXiv (opens in a new tab) - 08Agents16 Oct 2026
Perplexity open-sources SPACE, the sandbox behind its agents
Apache 2.0 sandbox platform behind Computer: each sandbox is a small VM with its own OS; credentials and network stay outside. Perplexity cites millions of sandbox creations a day, 3–5x faster than its previous provider.
Source: Perplexity (opens in a new tab) - 09Agents16 Oct 2026
Datadog Scout 1.0: AI SRE model claims frontier quality, 3-8x cheaper
Datadog's first incident-investigation model (built on GLM-5.3 Flash) matches GPT-6.1 Sol high and Claude Opus 5.5 high on its internal AlertEval benchmark at 3–8× lower estimated cost, per Datadog; in closed preview.
Source: Datadog (opens in a new tab) - 10Agents16 Oct 2026
Four Gemini 4 Argon agents beat Minecraft in 27 hours
In a new arXiv paper, four Gemini 4 Argon coding agents went from a random world to defeating the Ender Dragon in 27h 14m, writing 2,783 programs with no human input. Caveats: one run, Easy mode.
Source: arXiv (opens in a new tab) - 11Agents14 Oct 2026
Liquid AI releases LFM2.5-24B-A2B for on-device agents
Open-weight 24B MoE with ~2B active params and 131K context. Liquid reports LiveClaw 48.09 vs Qwen3.6-35B's 38.19, AutomationBench 27.69 vs 8.72, 184 tok/s on Apple M5 Max, and 21K+ output tok/s on one NVIDIA B300.
Source: Liquid AI (opens in a new tab) - 12Agents13 Oct 2026
Claude can run up to 1,000 agents in one workflow
Managed Agents workflows run up to 64 threads at once and 1,000 agents per run, so jobs too big for one conversation can fan out.
Source: Anthropic (opens in a new tab) - 13Agents13 Oct 2026
Z.ai launches GLM-Legal
An agentic legal model scoring 19.2% all-pass on Harvey LAB, 12.5 points above its base model, with results checked by Vals.
Source: Z.ai (opens in a new tab) - 14Agents13 Oct 2026
SubQ Snap makes agent decisions in about a second
Subquadratic's decision model picks from allowed answers instead of writing text, reads up to 384K tokens of text and images, and in Subquadratic's tests beats TypeSafe's Jev on 13 of 16 enterprise benchmarks. $0.05 per 1M input tokens, output free.
Source: Subquadratic (opens in a new tab) - 15Agents13 Oct 2026
AMD Helion: agentic AI racks up to 32,768 cores
Open OCP rack blueprint for agentic AI: CPU-only configs up to 32,768 EPYC cores; all-in-one with MI350P GPUs. Spec headed to OCP; partner configs planned November 2026.
Source: AMD (opens in a new tab) - 16Agents13 Oct 2026
AI agents got 48 hours to do real ML research
Hillclimb's DeepRSI: 15 open-ended ML problems, 48 hours, no internet. Claude Fable 5.1 leads at 1548 Elo ahead of GPT-6 Astra (1232); Astra even hard-coded a puzzle solver into a transformer's weights.
Source: Hillclimb (opens in a new tab) - 17Agents13 Oct 2026
DigitalOcean launches tools for self-improving agents
Agent Learning Services (Insights, Signals, Simulations) turn production sessions into tests before changes ship; Agent Data Services is a managed data layer for agent context. Learning in public preview, Data in private preview.
Source: DigitalOcean (opens in a new tab) - 18Agents13 Oct 2026
AgentSWE: coding agents fail to build the agents you want
25-task benchmark for creating/editing/optimizing agents. Strongest Creation mean 19.1; Editing 31.6; Optimization <5% gap closed on average. 11 Creation runs with ≥90 on development averaged only 9.9 held-out.
Source: arXiv (opens in a new tab) - 19Agents13 Oct 2026
Stateless AI agents cut long-horizon research tokens
SLA keeps research state in the harness and gives each agent a fresh context. On Anthropic kernel optimization with Codex, it matched the strongest baseline with 93.1% fewer tokens (67.9M vs 986.3M). Best full-budget result on every task; Advisor used 0.24–0.51% of tokens.
Source: arXiv (opens in a new tab) - 20Agents13 Oct 2026
ScaleDown CDMs: claims 200x faster tools in LangChain
LangChain wired ScaleDown's Summarize, Compress, Extract, and Classify CDMs into Deep Agents. ScaleDown reports CDMs are 200× faster and 467× cheaper than frontier models on these narrow tasks; up to 1M-token inputs; Summarize/Extract/Classify multimodal. On ScaleDown's benches, Summarize was 41× cheaper and Compress 138× cheaper vs an average of GPT-5.5 / 5.4 Mini / Nano.
Source: LangChain (opens in a new tab) - 21Agents13 Oct 2026
LightOn ColGrep-Agent-2B cuts coding-agent search tokens ~34%
LightOn's open 2B local code-search agent cut Claude Code API tokens by 34% vs grep on 52 code-search questions across 13 unseen repos (LightOn's test; not full coding tasks).
Source: LightOn / Hugging Face (opens in a new tab) - 22Agents13 Oct 2026
Anthropic rebuilt source control for AI agents
Merges grew from ~2,500/day in May to 30,000+/day by September, so a team of seven built Scion, an in-house VCS, and migrated off Git in five months. Anthropic reports 200x smaller infra, CI checkouts in seconds and a ~25% CI CPU win.
Source: Anthropic (opens in a new tab) - 23Agents12 Oct 2026
Google launches a Gemini agent for work
One agent for Gmail, Docs, Sheets, and code takes a goal and hands back finished work, in private preview.
Source: Google (opens in a new tab)