The Pulse — February 20, 2026

Google is explicitly framing “3.1 Pro” as the new baseline for complex reasoning + agentic workflows, shipping across API/Vertex/app/NotebookLM. They cite a big jump on ARC-AGI-2 (77.1% verified) and position it as “core intelligence” behind Deep Think.

Opus 4.6 adds 1M context (beta) and Anthropic is pushing: agent teams, compaction, adaptive thinking, and effort controls as product primitives — not just model claims. This shapes how “long-horizon agents” are built (memory/compaction + tool orchestration).

If this standard lands, it’s the endgame for “agent clicks buttons” UI automation. Websites can expose schema’d tools directly to an agent, which is both cheaper (tokens) and massively more reliable than vision + DOM scraping. Also: it changes the “AI browser” competitive surface (Chrome/Edge features, security model, consent, auditing).

Diffusion LMs keep coming back because they can parallelize, but they’ve been slow (lots of steps) and can’t KV-cache like AR LMs. CDLM claims 3.6×–14.5× lower latency while keeping quality on math/coding by (1) consistency training for multi-token finalization and (2) block-wise causal mask for KV cache compatibility.

Everyone is hitting KV cache walls with long context. This paper proposes latent-space KV compaction that’s far faster than prior “Cartridges”-style approaches, pushing “50× compaction in seconds” with little quality loss (per abstract). If true, it directly enables cheaper long-context agents and retrieval-heavy apps.

Small release notes can be the earliest signal of how “AI coding agents” will actually be used daily. Worktree isolation and background agents make long-running tasks safer and more parallelizable — i.e., closer to a real junior engineer workflow.

Creator-watch only matters if we go upstream: this episode points to concrete primitives (subagents, terminal UX, verbosity, teams) that show up as product changes elsewhere. It’s also a “how the sausage is made” reference for explaining agentic coding choices to a dev audience.

This is upstream from the usual “Groq/Cerebras” creator discourse: Taalas claims a new architecture merging storage+compute on-chip, aiming for order-of-magnitude latency/cost changes. Even if the first-gen is aggressively quantized, the product thesis is: ubiquitous AI = sub-millisecond inference + cheap hardware specialization.