AI Agents

Oct 8

On October 8, 2026, Paul Graham wrote that Amazon banning agents is the first opportunity he has seen since Amazon was founded for a startup to build an Amazon competitor. The post drew 399K views. Here is the argument, what Amazon has actually done, where the thesis is strong and where it breaks, and what an agent-friendly store would have to build.

Oct 7

Researchers from Cisco Foundation AI and Carnegie Mellon report that personal AI agents infer a user's wealth from email and profile data, then recommend more expensive flights, insurance and PhD programs, without being told to. Here is what the paper measured, what the caveats are, and what to do before giving an agent your inbox.

Oct 7

On October 6, 2026, Shane Mac, CEO of XMTP Labs, posted that an AI agent shared his personal bank balances in his company Slack, as him, and the post reached about 670,000 views. A Community Note on the post calls it fake engagement farming. Nothing has been independently confirmed. This post separates the claim from the dispute and extracts the permission lessons that hold either way.

Oct 7

Two launches on October 6, 2026 attack the same problem from opposite ends. Gumloop's Agent Browsers let agents click through sites that have no API, with a password vault the agent never sees. AgentMail's AgentID gives agents their own verified identity through standard OpenID Connect. Here is how each works, what they cost, and the security questions to ask.

Oct 7

On October 6, 2026, Brett Adcock, the founder of Figure, launched Hark Pro, a proactive personal AI assistant from his separate company Hark. It uses a model built for computer use, logs into sites through an encrypted vault, and is meant to find work for you instead of waiting for prompts. Here is what it is, the reported pricing, how it differs from Muse, Dots and Grok Bot, and the claims to verify.

Oct 6

Meta, SpaceXAI, Nous and Cursor are all converging on the same picture: you manage a team of bots, and one bot delegates to specialists. In the same week, a journalist said half her Muse use cases vanished because the browser stopped working, and Apple moved to gate full disk access. Our opinion: the next bottleneck for agents is not intelligence. It is permission.

Oct 6

Gamma launched Gamma 5 on October 6, 2026: a rebuilt platform with an agent that writes custom code and draws shapes, a freeform editor, thousands of templates and about 20 connectors. The stated goal is to remove the repetitive look of AI-made decks. Here is what changed, how to evaluate it, and what the launch does not tell you.

Oct 6

On October 6, 2026 Sierra unveiled the Personal Agent Protocol, an open standard for how consumers' AI agents interact with businesses, built with Meta, Genesys, Instinct, Rocket, Shopify, Stripe and Walmart. A v0.1 specification is planned for later this month.

Oct 4

A viral essay argues that every coding-agent memory plugin is the same RAG pipeline in different clothes, and that agents need a structured Markdown brain instead. We unpack the argument, the open-source Operator Memory plugin built on it, and the strongest objections from a 100-comment Hacker News thread.

Oct 4

On October 3, 2026, researcher Paulos Yibelo announced a full guest-to-host VM escape in KVM, and Vercel CEO Guillermo Rauch confirmed it through the Vercel Sandbox bounty program. There is no CVE, patch, or write-up yet. Here is what is verified, what is guesswork, and how to design agent sandboxes that survive a hypervisor bug.

Oct 3

On October 2, 2026, Cua shipped Cua Spaces — a free, source-available macOS app for running AI agents on desktops you can watch, take over, and reach across machines you own. Here is what Spaces are, how cross-computer access works, pricing, and what to set up this week if you already run computer-use agents.

Oct 3

On October 3, 2026, Shrivu Shankar's August essay "The Harness Is the Company" resurfaced on Hacker News with ~94 points. The claim: every SaaS becomes a harness around a model, and humans become taste-holders. This is not a reprint — it is a builder's own-vs-buy map, how the thesis differs from explainx.ai's agent-harness and software-factory coverage, and where the essay overreaches.

Oct 3

Perplexity says its Computer agent built a map of nearly 26,000 New York City restaurants and cafes where you can search by dish or neighborhood and step inside places like Peter Luger. We opened the app and read its own method page: the interesting part is how carefully the data is sourced, and how the headline numbers hold up.

Oct 2

On October 1, 2026 Earendil tagged Pi 1.0 and published experimental Pi Durable for long-running agent apps. This is not a changelog. It is how people actually use a minimal harness — tiny prompts, inspectable sessions, oh-my-pi's token tax, the fullscreen TUI fight, cache warming, and when a Slack or Kubernetes session beats another coding CLI.

Oct 1

On September 30, 2026, LiteLLM CTO Ishaan Jaffer launched Lens: an early-access analyzer that reads OpenTelemetry traces sent through the LiteLLM proxy, groups recurring agent failures, and lets teams query with SQL and APIs. Traces stay in customer ClickHouse; investigation results land in PostgreSQL. The 200,000-trace figure is a future scenario, not reported usage.

Sep 29

On September 22, 2026, Baselayer launched an Agentic Identity Suite — Know Your Agent (KYA) plus counterparty verification — for banks already running its business-KYB stack. explainx.ai maps what builders and FIs actually do with agent identity versus NVIDIA OpenShell sandboxes and agent wallets.

Sep 29

ElevenLabs shipped Eleven v4 and Eleven v4 Turbo on September 28, 2026 with a new expressive architecture, inline performance tags, Professional Voice Clone support restored from v3 gaps, and Turbo tuned for agent loops at ~100ms median inference. explainx.ai breaks down API model IDs, pricing promos, and when to pick Turbo over flagship v4 for realtime voice agents.

Sep 28

Most "agent memory" is RAG with extra steps: embed the chat, vector-search it, hope for the best. Hindsight, from Vectorize.io, takes a different bet — biomimetic memory types, a retain/recall/reflect loop, and observations that get refined rather than overwritten. 38,000 GitHub stars in, here's what it actually does, how to run it, and where the claims need a second look.

Sep 28

Manus announced Manus 2.0 on September 28, 2026 — not a patch release but a new architecture (Cascade), a renamed desktop app (Manus Studio) with timeline video editing and Game Dev, always-on Cloud Computers, and a separate Cue app for phone-first personal agents. Here is what shipped, what costs extra, and how it fits after Manus left Meta.

Sep 28

20 Claude Code tabs open and no idea which one is doing what — that's the problem Paperclip exists to solve. It's not another agent framework; it's the org chart, budget, and governance layer that sits on top of the agents you already have. 90,500 GitHub stars in, here's what it does, how to run it, and where the "not a chatbot" pitch is worth a second look.

Sep 27

Sep 25

Sep 24

A Stanford and Berkeley researcher released CLM, a System One model that scores actions by matching state and action embeddings. CLM-8B is on par with Jev on computer use, gaming and tool calling at up to 9x lower latency, and as a fine-tuned verifier it reaches 81.6 percent on DeepSWE and 87.6 percent on Terminal-Bench 2.1. We break down the charts, the small-sample caveats and how to try it.

Sep 23

Hosting an AI agent reliably — with proper isolation, tool access, and billing that doesn't punish idle time — is still real infrastructure work most teams would rather not own. DigitalOcean's new Managed Agents product targets exactly that gap: pay-per-use CPU billing and a 16,000-tool library, aimed at developers who want to deploy an agent without building and maintaining the hosting layer themselves.

Sep 23

Sep 23

As AI shopping agents increasingly navigate checkout flows directly, the amount of markup, styling, and boilerplate an agent has to parse before completing a purchase has become a real, measurable cost. Stripe upgraded 7.8 million of its own hosted checkout pages specifically to cut that token overhead by 42% — a concrete, quantified example of "agent-readiness" as an infrastructure investment, not just a buzzword.

Sep 23

"The right way to use model capabilities is not to ship 10x more features to prod" — Anthropic's Thariq Shihipar made a direct case against feature-velocity as the point of better models, arguing the actual leverage is in spending saved time on understanding users and prototyping, not output volume. The replies split between agreement and a pointed question: does this survive contact with a business that just wants more shipped?

Sep 23

Unreal Labs shipped an open-source Go agent harness that claims up to 40% cost savings over Codex on real benchmarks, by letting models issue tool calls asynchronously instead of blocking on each one. The name caused immediate confusion with Epic's Unreal Engine, and the benchmark methodology drew its own scrutiny — here's what the harness actually does, and what held up.

Sep 22

A tier-list image ranking agent harnesses — Oh My Pi alone in S-tier, Claude Code and Codex lumped into B alongside Cursor and Grok Build, GitHub Copilot and Antigravity in F — went viral on X September 21, 2026, racking up nearly 80,000 views and a comment section that disputed almost every placement. The single loudest complaint: Hermes doesn't appear on the list at all. Here's what the list actually claims, why the pushback matters more than the ranking, and how to build your own opinion instead of borrowing this one.

Sep 22

Sep 19

Sep 19

Sep 19

"AI evals" has become one of 2026's hottest practitioner skills — Hamel Husain and Shreya Shankar's course reportedly trained 2,000+ engineers and PMs, including teams at OpenAI and Anthropic. Here's what an eval suite actually is, the mistake most teams make first, and the scoring mix practitioners actually recommend.

Sep 18

Exa introduced Snapshot on September 18, 2026: a search API that returns results from the web as it existed on a date you specify, backed by more than 400 billion historical webpage snapshots spanning two decades. The pitch isn't browsing old pages for nostalgia — it's preventing web leakage in RL training and reproducible evals, and enabling point-in-time backtesting that previously took months of manual data collection.

Sep 18

Google Labs introduced CC on September 18, 2026 — an AI agent built around household logistics rather than individual productivity: a shared "Your Day Ahead" morning brief, autosynced Google Calendar and Tasks across up to five family members, meal-plan drafting in Google Chat, and delegated paperwork like school permission slips. It's US-only, 18+, and waitlist-gated.

Sep 18

Claude Code, pi, and Hermes look different on the surface but solve the same ten underlying problems. These are the concepts that separate a demo that falls over after ten turns from an agent you can trust to run unattended — each with a concrete example and a way to build it yourself.

Sep 18

Claude Code, pi, and Hermes all call the same model APIs. What separates a working coding agent from a demo that falls over after ten turns is everything wrapped around the model: the agent loop, the tool contracts, the context and memory system, and the recovery logic that keeps a session alive across failures. That layer now has a name — harness engineering. Here's what it actually covers.

Sep 17

Monid open-sourced a tool router that gives AI agents standardized access to 2,000 different APIs through a single integration layer — targeting one of the more persistent practical problems in agentic AI development: every new tool an agent needs typically requires its own custom integration work.

Sep 16

Elon Musk's tweet about Blizzard's September 15, 2026 login meltdown cited E.M. Forster's 1909 story "The Machine Stops" — the classic warning about what happens when the people who understood a system are gone and only the automation is left. That's not really a gaming story. It's a direct warning for any team running AI agents in production without keeping the human understanding underneath them.

Sep 16

Plasma AI launched Radio on September 14, 2026 — a shared channel any agent "that can fetch a URL" can join and use to message other agents and humans directly. It's a much simpler mechanism than it sounds: no new protocol, just a URL-based relay. Here's what it does, what it doesn't, and how it compares to MCP.

Sep 16

Sep 15

Sep 15

Sep 15

Sep 15

Sep 13

Sep 13

Sep 13

Sep 11

Sep 11

Sep 11

Sep 10

Sep 9

Sep 8

Sep 7

Sep 6

Sep 5

Sep 5

Sep 5

Sep 5

Sep 4

Sep 3

Sep 3

Sep 1

Sep 1

Sep 1

Sep 1

Sep 1

Aug 31

Aug 31

Aug 30

Aug 29

Aug 29

Aug 29

Aug 29

Aug 29

Aug 28

Aug 27

Aug 26

Aug 26

Aug 25

Aug 25

Aug 22

Aug 22

Aug 22

Aug 21

Aug 20

Aug 20

Aug 20

Aug 20

Aug 20

Aug 19

Aug 18

Aug 18

Aug 18

Aug 17

Aug 17

Aug 15

Aug 15

Aug 14

Aug 12

Aug 12

Aug 11

Aug 11

Aug 11

Aug 10

Aug 10

Aug 10

Aug 8

Aug 6

Aug 5

Aug 5

Aug 5

Aug 5

Aug 3

Aug 3

Aug 2

Aug 1

Jul 26

Jul 26

Jul 26

Jul 26

Jul 25

Jul 24

Jul 22

Jul 22

Jul 22

Jul 22

Jul 21

Jul 21

Jul 20

Jul 18

Jul 17

Jul 16

Jul 14

Jul 12

Jul 9

Jul 3

Jul 3

Jun 30

Jun 30

Jun 29

Jun 29

Jun 29

Jun 28

Jun 28

Jun 28

Jun 28

Jun 28

Jun 28

Jun 28

Jun 28

Jun 27

Jun 27

Jun 27

Jun 25

Jun 24

Jun 24

Jun 24

Jun 24

Jun 24

Jun 24

Jun 23

Jun 23

Jun 23

Jun 23

Jun 22

Jun 21

Jun 20

Jun 20

Jun 19

Jun 19

Jun 19

Jun 17

Jun 17

Jun 16

Jun 16

Jun 16

Jun 16

Jun 14

Jun 14

Jun 13

Jun 12

Jun 10

Jun 9

Jun 9

Jun 9

Jun 6

Jun 4

Jun 1

May 31

May 30

May 30

May 23

May 22

May 22

May 22

May 20

May 20

May 17

May 14

May 14

May 13

May 12

May 9

May 8

May 8

May 8

May 7

May 7

May 6

May 4

May 4

May 2

Apr 24

Apr 23

Apr 22

Apr 13