Signals from the agent world

Visual Edits adds a click-and-change editing layer to Bolt.new: clicking any element in the running app opens a toolbar for in-place text, font size, styling, lists, links, and color changes, with the color picker reading the page's existing palette. Changes stack into a list and go to the model as one instruction on save — exact values instead of a description it has to interpret. Dragging and layout rearrangement are not supported; all free and paid users get it.

Warp launches Factory Benchmarks in early access, letting teams evaluate model and harness configurations on their own coding tasks rather than public benchmarks like SWEBench. Benchmarks defined via factory.yaml replay prior runs across a configuration matrix; Scorers grade runs with an LLM-as-a-judge loop; Pareto-style views map dimensions like cost, quality, correctness, and verbosity. Warp reports about 63% lower internal cost per PR on some task types.

Campaign Composer extends Decagon's Proactive Agents into outbound campaigns over voice calls and SMS, self-serve configured: running hours, retry logic, attempt caps, and cadence, with contact-field suppression, opt-out and wrong-party flags, and built-in PII/PHI redaction for regulated programs such as debt collection, prescription refill reminders, and travel upgrades. Campaigns follow up over days or weeks within set limits; Hertz is cited as a customer.

Enterprise Frontier Safeguards pairs zero-data-retention privacy with misuse detection: data stays in customer-controlled cloud storage while automated monitoring scans rolling traffic windows for serious misuse — including agents engaging in destructive behavior. Built with AWS, Google Cloud and Azure plus input from 100+ customers, it supports surfaces including Claude Code, Claude Enterprise, Bedrock and Microsoft Foundry, free, rolling out in phases this fall.

Anthropic released Fable 5.1, generally available on AWS, Google Cloud, Azure and the Claude apps, and Mythos 5.1 — the same model under stricter access controls, limited to vetted organizations via its Cyber and Life Sciences Verification Programs. Terminal-Bench 4.0 reaches 60.9% for Mythos 5.1 vs Fable 5's 42.0%, with pricing unchanged at $10/$50 per million tokens and cache reads down 75%.

Clay Sequencer is now generally available on Launch, Growth, and Enterprise. Built on Clay's Audiences data layer, it unifies lead sourcing, enrichment, and cold email sequencing: segments continuously refresh from CRM data, signals, or Clay's contact database, eligibility is verified at send time, and agents fill templated fields from 200+ providers with required-field guardrails, plus deliverability tooling for domains, inbox warming, and bounce analytics.

Horizon Scanning lets legal teams monitor regulatory change by describing what to track in plain language — a topic, region, or named regulator — drawing on 12,000+ sources across 100+ jurisdictions. It surfaces the most consequential updates first in one dashboard, with relevance assessments for gauging urgency, sends daily email digests, and connects alerts to downstream work such as memo drafting and client alerts. Early Access for Harvey customers; GA to follow.

Elicit launches four collaboration features. Skills let researchers formalize workflows and share them across the organization; Collaborative Sessions put multiple researchers in one session with shared context and version history; Artifact Editing lets users direct the Research Agent to extend prior work or open side threads; Shared Projects give teams a persistent space for related research. Skills also ship on Pro; the rest need Scale or Enterprise.

Augment Code launched Cosmos Advisor, an AI forward-deployed engineer embedded in its Cosmos platform. From one conversation it answers how the platform works, configures triggers and environments, deploys bespoke agents, and diagnoses failed runs — within the platform's object-level access controls. Augment cites its own on-call use: an Incident Investigator agent cut human investigation work by 81.3% and median time to first RCA from 30.1 to 6.2 minutes.

Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to safely operate physical devices — lab instruments, robotic arms, quantum-computer lasers — via MCP, command line or code APIs, with standard-format device discovery and enforced safety limits. Case-study partners include Genentech, Carnegie Mellon, HHMI Janelia and QuEra, with AWS, Danaher, QIAGEN and Universal Robots among the vendors building support.

In open beta, teams can drop an Account Agent into any Clay Workflow as a node. The agent arrives knowing the account's full history — prior contacts, past campaigns, deal history, and previous conclusions — and decides which action to take from the set defined in Workflows. Outputs flow into branching, enrichment, or Slack notifications; findings are written back to Audiences so scoring self-improves over time. Available on Enterprise, Growth, and Launch plans.

Sierra announces the opening of its Seoul office, a deliberate step into the Korean market for sophisticated, enterprise-grade customer-experience agents. Sierra describes the launch as part of a wider international expansion that pairs local presence with its voice personas, context engine, and agent guardrails for the most demanding brands.

Harvey Tenet is Harvey's first post-trained open-weight model, built on a Kimi K3 base and co-trained with Fireworks research for long-horizon legal work. It launches as a research preview on the Harvey platform, where it powers the legal reasoning layer of Harvey II.

Decagon Assist is an AI copilot for the human representatives on a team. It surfaces context from earlier AI-agent conversations, suggests next-step actions, drafts brand-aligned responses, answers knowledge-base queries, transcribes live voice with guidance, and translates chats in real time — all inside existing systems like Salesforce, Zendesk, and Front, sharing a knowledge base with Decagon's agents.

Clay Workflows is the company's new orchestration primitive: a visual graph of connected steps where a trigger starts the flow, each record moves through branches one at a time, and results write back to Audiences. Workflows can be built headlessly through the Clay CLI or conversationally via Sculptor; Clay's growth team replaced an 81-column table with a single workflow for product-led vs. sales-led routing, dedupe, outreach, and self-healing error monitoring.

Harvey II is a major platform update: agents now open inside Spaces with the matter's documents, parties, tasks, permissions, and history already loaded. Persistent memory learns each user's style, structure, citations, and tone across Harvey, Word, and Outlook. The release also debuts Harvey Tenet — Harvey's first post-trained open-weight model for long-horizon legal work (Kimi K3 base, co-trained with Fireworks).

Warp Factories enters closed beta: open infrastructure for cloud software factories — fleets of coding agents automating the SDLC from triage through verification. A foreman orchestrator routes work from tools like Slack and Jira, spawning subagents with chosen models and harnesses spanning Warp's agent, Claude Code, Codex, and open-weight models. Factories are version-controlled code with bring-your-own inference; Warp says about 30% of its own tasks run this way.

Genspark open-sourced GenOffice, which it calls the world's first full-featured open-source AI office suite and positions as an AI-native alternative to Microsoft Office: free, ad-free Docs, Sheets, Slides and PDF editors for PC and Mac, released under Apache-2.0.

Anthropic ships a step change for the Opus tier, tuned for long-running agents and heavy professional work.

OpenAI introduced Presence, a managed product for deploying trusted AI agents across enterprise voice and chat workflows. Agents ship with approved actions, guardrails, and human escalation, can be tested in simulations before going live, and improve through a Codex-driven loop. OpenAI describes it as a battle-tested product that answers questions and resolves issues for enterprises.

Anthropic's new Sonnet delivers frontier performance across coding, agents, and professional work at scale.

Agentic presentation generation — outline, design and export a deck in one run.

Developers can embed v0's generative UI pipeline inside their own tools.

An agentic browser where the assistant can act on the page you're looking at.

Arc's successor puts a chat-with-your-tabs agent at the center of the browser.

Anthropic takes its terminal coding agent out of research preview alongside the Claude 4 launch.

GitHub assigns issues to an autonomous agent that works in Actions-powered environments.

Brief in, brand system out — the first 'design agent' positions itself between Midjourney and Figma.

A mixture-of-agents system that can make phone calls and generate Sparkpages on demand.

The invite-only general agent demo spreads across social feeds; the waitlist crosses one million within weeks.

The long-horizon research agent that kicked off an entire category — and a naming convention.

A research preview of a browser-using agent, powered by the Computer-Using Agent model.

The first widely publicized 'AI software engineer' demo resets expectations for autonomous coding.

The research preview that reached 100M users in two months — and started the whole map.