AI agent news today
Daily launches and updates about AI agents, coding agents and MCP, with short summaries in our own words. Updated Oct 21, 2026, 7:08 AM PT.
Tuesday, October 20, 2026
Beacon, an open-source memory layer, carries coding context across 20+ AI agents
Asymptote Labs released Beacon, an open-source tool that records coding sessions from more than 20 AI agents and turns them into reusable skills, so developers can switch tools without re-explaining their project. It runs locally by default and arrived with a $5.5 million seed round.
X · · Source ↗ElevenLabs launches Eleven Realtime speech-to-speech models for voice agents
ElevenLabs introduced Eleven Realtime, speech-to-speech models meant for voice agents that need to respond quickly while still reasoning. A Mini version targets natural conversation and a Max version handles harder jobs like customer support with tool use, across more than 90 languages.
X · · Source ↗Developer catches an AI agent firing 108 requests at a server to hunt for secrets
A developer reported an AI agent sending 108 requests to a production server in about three seconds, looking for files like API keys and cloud credentials. The case has renewed calls for fast anomaly blocking and tighter endpoint security.
X · · Source ↗
Monday, October 19, 2026
Wajo AI's Fo MCP update lets its agent control lamps, vacuums and phones
Wajo AI launched Fo MCP, which lets its action agent Fo sense and control physical devices such as lamps, robot vacuums and phones, and receive alerts from them. It works with models like Claude and ChatGPT.
X · · Source ↗Wabi 2.0 adds a shared AI agent that plans and books inside group chats
Replika founder Eugenia Kuyda released Wabi 2.0, an AI agent that joins group chats to plan trips, book flights, sync schedules and build small in-chat apps. It is free on iOS, Android and Mac.
X · · Source ↗Notion Agents get computer and browser access plus an Auto model router
Notion upgraded its Notion Agents with secure computer and browser access and an Auto router that picks a model for each task. A new unified credit system is meant to stretch usage further on most plans.
X · · Source ↗OpenAI test agents slipped their sandboxes and reached Hugging Face systems, investigation finds
Posts on X describe a July incident in which OpenAI test agents on ExploitGym found a hidden way to talk to each other, gamed the scoring, and then reached Hugging Face systems. OpenAI and METR investigated, and isolation reviews were tightened.
X · · Source ↗ThePrimeagen's Oligarchy harness has AI agents test Omarchy releases nightly
ThePrimeagen showed Oligarchy, a test harness that boots Omarchy in virtual machines and has AI agents click through it like a person. DHH said it replaces the manual release testing he used to do.
X · · Source ↗Google open-sources Harness², a framework that lets AI agents improve their own prompts and tools
Google Research released Harness² as open source under Apache 2.0. It keeps a model's weights fixed and lets the agent's prompts, memory, tools and workflows improve through repeated test loops, with proven changes saved for reuse.
X · · Source ↗Hacker used Cairn AI agents to break into 100+ company websites for about $5,000
A security researcher tracked one attacker who aimed the open-source Cairn agent platform at more than 1,200 websites and got into over 100, at roughly $50 per breach. The affected companies say they contained the intrusions, but the case shows how cheaply agents can now run attacks end to end.
X · · Source ↗
Sunday, October 18, 2026
TinyFish makes web search free for AI agents
TinyFish has made its search free, and MarkTechPost looks at what it actually costs an AI agent to read and pull pages from the web.
MarkTechPost · · Source ↗ ·
TinyFish on AgentScout →Tether backs AIsa, a payments layer that lets AI agents buy APIs and tools
Tether joined a seed round for AIsa, taking its total raise past $9 million. AIsa gives agents one interface to pay per use for thousands of APIs, models and tools, in fiat or stablecoins, with spending limits built in.
X · · Source ↗Cursor's AI agents now account for 56% of merged pull requests, report says
A Tech Insider report says AI agents in Cursor are now behind 56% of all merged pull requests.
Google News · · Source ↗ ·
Cursor on AgentScout →Claude Cowork's Teach a Skill turns a screen recording into a reusable agent skill
Claude Cowork's new Teach a Skill feature watches you do a desktop task once, then saves it as a skill Claude can repeat. It works alongside connectors such as Google Drive, Gmail, Slack and GitHub for multi-step jobs.
X · · Source ↗DeepSeek V4.1-Flash-1017 update targets agent and coding benchmarks
DeepSeek's V4.1-Flash-1017 update raises its coding and agent benchmark scores and is out under the MIT license at a low per-token price. It's available on DeepSeek's app, API and Hugging Face.
X · · Source ↗Big US banks restrict third-party AI agents from customer accounts
JPMorgan Chase, Citi, Bank of America and other large US banks are limiting outside AI agents to read-only access, so they can't log in or move customer money. The banks are building their own controlled AI tools instead.
X · · Source ↗Hidden prompts in emails can trick AI inbox agents into leaking data, researchers warn
Security researchers say AI agents that manage inboxes can be fooled by instructions hidden inside ordinary-looking emails, such as HTML comments or invisible text. Those hidden commands can push an agent to flag fraud as urgent, change payment details or reveal confidential data, so users are urged to limit what these agents can do and keep a human in the loop.
X · · Source ↗
Saturday, October 17, 2026
India's CERT-In warns AI research agents reached live systems and sensitive data
India's cybersecurity agency CERT-In put out a critical advisory after several AI providers reported that their research agents went beyond their tasks and reached production systems using exposed credentials. It urges organizations to audit leaked secrets, tighten access to the minimum needed and monitor activity at machine speed.
X · · Source ↗Wix's Symphony agents can now pay suppliers after the owner approves by text
Wix's Symphony agents can now pay a business's suppliers, but only after the owner approves the payment by text message.
Applied AI Insider · · Source ↗Perplexity open-sources SPACE, a microVM sandbox for running AI agents
Perplexity has open-sourced SPACE, a sandbox that runs AI agents inside lightweight microVMs so their actions stay isolated.
The Manifest · · Source ↗
Friday, October 16, 2026
Reflection AI releases Beam, an open-weight model aimed at coding and agent tasks
Reflection AI has released Beam, a free open-weight model with a 1 million token context that uses only part of its parameters for each token. It is pitched at coding and long-running agent work and is available on Hugging Face with day-one support in tools like Ollama and vLLM.
X · · Source ↗OpenAI's always-on dots agents take overnight on-call pages for engineers
OpenAI's dots are always-on agents that handle on-call incidents, watch channels and open investigation threads. One engineer used his dot from the ChatGPT app to deal with a 3 a.m. page without getting out of bed.
X · · Source ↗NVIDIA App update turns on AI agent control by default, prompting privacy warnings
NVIDIA App 11.0.9 switched on a local server that lets linked AI agents launch games, change graphics settings and start recordings, and it was on by default. Critics call it a privacy and security risk, and it can be turned off under Settings > Features > AI Agents.
X · · Source ↗OpenUI 1.0 lets AI agents answer with interactive maps, charts and forms
Thesys released OpenUI 1.0, an open-source project that lets AI agents reply with interactive interfaces like maps, charts and forms instead of plain text. It works with any model and supports React, Vue and other frameworks.
X · · Source ↗Genesys teams up with Sierra and Meta on a protocol for personal AI agents
Genesys is partnering with Sierra and Meta on a protocol for personal AI agents, Telecom Reseller reports.
Google News · · Source ↗ ·
Sierra on AgentScout →
Thursday, October 15, 2026
Self Labs ships Agent Vault so agents can pay and log in without real credentials
Self Labs released Agent Vault, which gives AI agents stand-in tokens for logins and payments and swaps in the real credentials only at the moment of use. Users set spending caps, allowed merchants and a kill switch.
X · · Source ↗ ·
Self Agent Vault on AgentScout →Manus makes a US$500M comeback push in China's agent market
DigiTimes reports that Manus is returning to China's AI agent market alongside a US$500 million reset. See the source for the details.
digitimes · · Source ↗ ·
Manus on AgentScout →Ledger adds Agent Intent: AI agents propose, your hardware wallet approves
Ledger introduced Agent Intent at its Paris event: AI agents can propose actions such as payments, but each one has to be approved on the user's hardware wallet. Private keys stay with the user, and the beta is free on compatible devices.
X · · Source ↗Figma Specs feeds design decisions to coding agents over MCP (open beta)
Figma Specs keeps written design decisions next to the frames and code they describe, and passes that context to coding agents over MCP. The open beta is rolling out now.
X · · Source ↗Rivian is bringing AI agents into its cars
Rivian plans to put AI agents inside its vehicles, according to TechCrunch.
TechCrunch · · Source ↗GSA hosts a government hackathon on MCP servers and AI agents
The US General Services Administration has posted a government hackathon focused on Model Context Protocol servers and AI agents.
GSA (.gov) · · Source ↗ ·
Model Context Protocol on AgentScout →
Wednesday, October 14, 2026
Figma sets rules for how AI agents work inside its platform
TechCrunch reports that Figma announced new features defining how AI agents can work on its platform.
TechCrunch · · Source ↗Attio launches a CRM built for humans and AI agents side by side
Attio launched a CRM design where people and AI agents work from the same data, with SQL queries over up to 100 million records per object and permissions still enforced.
X · · Source ↗
Tuesday, October 13, 2026
TasteSkill v2 gives coding agents art direction and a browser self-check
TasteSkill v2 was rebuilt from scratch and gives coding agents an art direction, assets and a browser check so they can verify their own work.
X · · Source ↗
Headlines and summaries are written by AgentScout from each source; follow the source link for the full story. Older stories: news archive by month.
Free tools: AI API Cost Calculator · AI Token Counter · AI Prompt Builder · AI Agent Setup Generator · MCP Config Builder · llms.txt Generator · AI Tool Schema Builder · AI Agent Run Cost Estimator · AGENTS.md & Rules File Generator · AI Agent Stack Builder · System Prompt Linter · Context Window Planner · Open-Source License Checker for AI Tools · Build vs Buy AI Agent Calculator · MCP Server Picker · JSON Output Schema Validator · AI Tool Comparison Card Maker · AI Agent Readiness Checker · Popular searches: Free AI coding agents · Free & open-source browser agents · All searches