Anthropic Leads AI Roundup 2026 as Autonomous Agents Take the Wheel in February

Anthropic’s “Projects” platform, which lets teams spin up shared Claude workspaces, saw a surge in adoption after the company disclosed a coordinated distillation attack on its Claude model. In a self‑published report, Anthropic named three Chinese labs—DeepSeek, Moonshot and MiniMax—as the perpetrators of a campaign that generated 16 million exchanges across 24 000 fraudulent accounts, then used those interactions to train rival models (Jacob Stephen, Feb 2026). The disclosure forced frontier labs to embed rate‑limiting, fingerprinting and behavioral analytics into their APIs, a defensive layer that had previously been optional. According to VentureBeat, the new “Projects” UI now includes granular sharing controls and audit logs, enabling enterprises to monitor who accesses which Claude instance and to enforce usage policies in real time (VentureBeat, 2024).

The shift from “assistant‑in‑the‑tool” to autonomous agents operating in dedicated environments accelerated in February. Cursor’s Cloud Agents launched a workflow where a task triggers a full‑stack VM that writes, tests and records a demo video before returning a merge‑ready pull request, eliminating the need for developers to review diffs (Jacob Stephen, Feb 2026). Perplexity’s “Computer” followed a similar model, routing sub‑tasks to specialized models for research, design and deployment without further prompting. OpenAI’s Codex app took a parallel‑agent approach, presenting a command center that orchestrates multiple coding agents simultaneously. All three products converge on the same premise: the agent becomes the execution environment, not merely a query interface.

Cheaper, continuously‑running models from Chinese providers dramatically altered the economics of agent deployment. Z.ai’s GLM‑5, Alibaba’s Qwen 3.5 and MiniMax M2.5 entered the market within a two‑week window, each delivering frontier‑level coding performance at roughly $1 per hour of continuous compute (Jacob Stephen, Feb 2026). The Kimi Claw and MaxClaw frameworks further lowered the barrier to entry by allowing agents to launch from a browser or embed directly into Slack, Discord and Telegram, respectively, without any server‑side setup (Jacob Stephen, Feb 2026). This price point encourages product teams to design services around perpetual automation rather than intermittent API calls, a trend that analysts predict will reshape SaaS business models over the next 12‑18 months.

The Pentagon’s $200 million defense contract request added a geopolitical dimension to the emerging agent ecosystem. Anthropic refused to certify its models for “all lawful purposes,” drawing a hard line that sparked a public dispute with the U.S. Department of Defense (Jacob Stephen, Feb 2026). While the contract remains unresolved, the episode underscores the growing regulatory friction as governments seek to harness autonomous AI while vendors grapple with export controls and ethical use clauses.

Legal pressures intensified alongside technical battles. A recent ruling by a federal judge granted Anthropic a partial victory in a lawsuit over AI‑generated book training data, affirming the company’s right to use publicly available text for model improvement (The Register, 2024). The decision, though narrow, bolsters Anthropic’s stance that large‑scale data ingestion is a legitimate practice, even as competitors face increasing scrutiny over data provenance.

Collectively, these developments mark February 2026 as a watershed moment: autonomous agents have moved from experimental add‑ons to self‑contained workhorses, cheap frontier models have democratized continuous AI operation, and both industry and governments are scrambling to adapt to the new security, economic and regulatory realities. Anthropic’s leadership in publishing the distillation‑attack report and expanding its collaborative “Projects” suite positions it at the forefront of this rapidly evolving landscape.