Radar Trends to Watch: February 2026

Loukides — Radar Trends to Watch: February 2026

The second installment in the wiki of Loukides’s monthly digest, published 3 February 2026, covering January 2026. Opening framing: “If you wanted any evidence that AI had colonized just about every aspect of computing, this month’s Trends would be all you need.” Loukides points to Steve Yegge’s Gas Town as the exemplar — a vision of multi-agent orchestration that he uses to anchor the month.

The month’s structural story is AI becoming infrastructure: tools that used to be model-distinct (Cursor’s hundreds-of-agents experiment), capabilities that used to require fine-tuning (Kimi K2.5’s 100-subagent swarms), and primitives that used to be in-scope for a single developer (NanoLang as a programming language designed specifically for LLM code generation) are all converging toward an AI-as-substrate default.

TL;DR — the month’s standout threads

• AI has colonized everything. Programming, security, operations, design, physical systems — all show substantive AI integration this month.
• Cursor’s hundreds-of-agents experiment — built a web browser in one week with parallel agent dispatch. Named as the existence-proof for at-scale multi-agent dev work.
• Kimi K2.5 enables swarms of 100 subagents out of the box; Qwen-3-Max-Thinking claims parity with Opus 4.5 and Gemini 3.
• MCP Apps becomes an official MCP extension for UI components — the wiki’s first source on MCP-as-UI-substrate, not just MCP-as-tool-substrate.
• Moltbook social network for agents — agents share thoughts; users can observe. First wiki source on agent-to-agent social infrastructure as a discrete category.
• OpenClaw publishes — provides LLMs persistent memory for always-on local agent operation. (March digest will name this as the month’s compound story.)
• Anthropic publishes the Claude constitution — detailed behavioral training framework as a transparency document.
• Kent Beck reframes AI’s effect on engineering — AI augments junior developers and accelerates learning cycles, rather than replacing them. Communications and soft skills emerge as the engineer’s most important capabilities in this digest’s editorial.
• A new ChatGPT exfiltration attack — leaks private user info without machine traces. Fine-tuning on bad code tasks causes misbehavior spillover to other contexts (model-quality / safety inversion, named as a fundamental finding).

Sectional summary

AI. Genie 3 (Google’s interactive 3D video generation world model); Kimi K2.5 (multimodal, 100-subagent swarms); Qwen-3-Max-Thinking; MCP Apps spec; Moltbook (agent-only social network); OpenClaw (always-on persistent agent memory); FlashWorld (text/image → 3D scene in seconds); knowledge bases benefit from negative examples and decision trees; Anthropic publishes the Claude constitution; OpenAI experiments with ChatGPT ads and ChatGPT Go.

Programming. Cursor’s hundreds-of-agents-built-a-browser experiment; AI-assisted programming relocates rather than eliminates rigor; Ghostty welcomes AI tool use but requires qualified human operators; NanoLang (language designed for LLM codegen); Unix bash tools as viable agent foundation vs specialized frameworks; Gleam on Erlang BEAM; Steve Yegge’s Gas Town as multi-agent orchestration vision; code-review methodology requires knowledge of prompts and human edits; Kent Beck on AI augmenting junior devs.

Security and Privacy. ChatGPT exfiltration attack (no machine traces); sandboxes for AI limit damage potential of misbehaving generated code; Google Search AI Mode accesses photos/email (training excludes personal data); fine-tuning on bad code causes misbehavior spillover to other contexts; California’s DROP privacy law (centralised consumer data-deletion requests); SSL cert management automation risks; MongoBleed; phishing trends 2025; Google shuts down dark-web report feature; Microsoft retires RC4.

Operations. AI stress-tests business models (operational value moves up the infrastructure stack); DNS resilience strategies; Kubernetes 1.35 (vertical scaling without Pod restarts).

Things. Google revisits smart glasses (Meta Ray-Ban early success); NVIDIA Vera Rubin GPU series (5× efficiency); AI vending machine tricked into giving stock and ordering unauthorized items; DeepMind builds automated materials-science lab integrated with Gemini.

Design. Design discipline often excluded from AI integration; design projects now deliver assets and AI tools for clients; effective design requires understanding both user and developer frustrations.

Cross-positioning with the wiki

• OpenClaw’s first wiki mention — will recur in March and later digests. Worth tracking as a candidate concept page once a third independent source treats persistent agent memory as a discrete primitive.
• MCP Apps spec complements the wiki’s agent-harness coverage at the agent-UI-surface layer (vs the tool-surface layer already documented).
• Kent Beck’s “AI augments junior devs” claim is a counterpoint to Brynjolfsson’s canary-in-the-coal-mine empirical anchor on early-career displacement. Worth a body-link on ai-employment-effects as a practitioner-side contradiction note.

Named entities (this ingest)

• Mike Loukides — see January.
• O’Reilly Media — see January.

Plus dangling first-mentions of: Steve Yegge, Kent Beck, Genie 3, Kimi K2.5, Qwen-3-Max-Thinking, MCP Apps, Moltbook, OpenClaw, Anthropic Claude constitution, NanoLang, Gleam, Ghostty, Gas Town, DeepMind, Vera Rubin.

Source-quality notes

Same as January: single secondary-summary source; confidence 0.65.