February was the month things moved from interesting to consequential. Agents got their own environments to operate in, the economics of running them flipped, and the wider world started pushing back on all of it.
Agents Got Computers
Cursor, Perplexity, and OpenAI all shipped independently this month and all landed at the same conclusion. Agents shouldn't just assist you from inside your existing tools. They should have their own environment, their own context, and their own way of showing you what they did.
Cursor Cloud Agents are the clearest version of this. You assign a task, the agent spins up in its own cloud VM, writes the code, tests it, and comes back with a video demo and a merge ready PR. You're not reviewing a diff. You're watching the feature run. A meaningful chunk of Cursor's own merged PRs now come from these agents.
Perplexity had been quiet for a while before dropping Perplexity Computer, a general purpose digital worker that routes tasks across a fleet of models, research to one, design to another, deployment to another, running through long workflows without needing to be prompted again at each step. OpenAI's Codex app takes a different angle on the same idea. Rather than one agent handling everything end to end, it's a command center where you spin up multiple coding agents in parallel, hand each one a separate task, and manage all their work in one place. The approach differs but the underlying shift is the same: the agent is no longer a tool you reach for. It's where the work lives.
The Economics Flipped
A wave of models from Chinese labs hit this month and the story isn't just more competition. It's that near frontier capability is now cheap enough to run constantly.
Think of it like midrange smartphones the moment they stopped feeling like compromises. GLM-5 from Z.ai, Qwen 3.5 from Alibaba, and MiniMax M2.5 all landed within the same two week stretch, each competitive on real coding tasks, each priced aggressively compared to many Western APIs. MiniMax runs at around $1/hour continuous. And both Kimi Claw and MaxClaw launched as agent frameworks that deploy in seconds with no server setup. Kimi Claw runs from your browser, MaxClaw embeds directly into Telegram, Slack, and Discord so the agent lives where your work already happens.
When capable models are this cheap to run constantly, you stop minimizing AI calls and start designing products built around continuous automation. That's a different product than most teams are building today.
Consequences Arrived
February was the first month where the wider world started pushing back, and it happened on three fronts.
Anthropic published a report naming DeepSeek, Moonshot, and MiniMax for running coordinated distillation attacks on Claude, generating 16 million exchanges with Claude across 24,000 fraudulent accounts and training their own models on those outputs. Distillation between your own models is standard practice. Doing it on a competitor's at that scale is something else. The community split on whether Anthropic's framing was fair, but the practical upshot is clear. Rate limiting, fingerprinting, and account behavior detection are now part of the competitive stack every frontier lab has to build.
Then, late in the month, a different kind of conflict came to a head. The Pentagon wanted Anthropic to allow its models to be used for all lawful purposes as part of a $200 million defense contract. Anthropic refused, drawing two hard lines: no fully autonomous weapons and no U.S. domestic surveillance. When negotiations broke down, the Defense Secretary designated Anthropic a supply chain risk and President Trump directed federal agencies to immediately stop using Anthropic’s technology. Within hours, Sam Altman said OpenAI had reached an agreement with the Pentagon to deploy its models on classified networks, and that OpenAI’s safety red lines on domestic surveillance and autonomous weapons were included in the deal. Anthropic said it intends to challenge the supply chain risk designation in court. As a signal of how entangled AI and national security policy have become, it's one of the more significant moments the field has seen.
Earlier in the month, Seedance 2.0 from ByteDance crossed a different kind of line. One of the most reliable tells that made AI video easy to spot was the mouth. Seedance takes direct aim at it with joint audio-video generation that bakes lip sync into the process rather than adding it afterward. A creator used it to generate a hyperrealistic fight scene between Tom Cruise and Brad Pitt from a two line prompt. The clip spread before most people stopped to think about what they were watching. The Motion Picture Association sent its first cease-and-desist letter to a major AI firm. Major studios followed. SAG-AFTRA called it blatant infringement. For developers building in this space, watermarking, provenance, and moderation are no longer optional. They're what keeps you in business.
The Pattern Behind It All
February was the month the gap between what AI can do and what the world is ready for started showing.
- Agents got computers and the trend is clear that agents are moving out of the sidebar and into dedicated environments, from cloud VMs to agent workspaces
- The economics flipped and cheap capable models change not just what you spend, but what kind of products are worth building
- Consequences arrived and the distillation report, Seedance, and the Pentagon standoff made clear the field is no longer operating without friction
The tools have taken the wheel. The question now is where we steer.
Top comments (0)