The Pulse — March 28, 2026
The signals that entered our radar, organized with sources and context to understand what changed.
The audio script is ready; narration will appear after voice generation finishes.
Put Claude to work on your computer (Dispatch + “computer use”)
WHY IT ENTERED THE RADARAnthropic is productizing “agent runs on your real desktop” (mouse/keyboard/browser) with a permissions + safeguard story. This is the most direct path from chat → doing (and it competes with third‑party agent frameworks).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Computer-use agents are here: what they can/can’t do, and the 3 failure modes (permissions, prompt injection, and ‘oops I clicked the wrong thing’).”
Claude Code “Auto mode” (permission automation with a safety classifier)
WHY IT ENTERED THE RADARThis is a concrete pattern we’ll likely see everywhere: delegate approvals to a policy/classifier layer instead of using YOLO flags. It’s a blueprint for safer long-running coding agents.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Auto-mode is the future of agent ops: how to think about ‘guardrails that decide’ vs ‘humans approve everything’.”
Gemini 3.1 Flash Live (real-time audio model + benchmarks)
WHY IT ENTERED THE RADARGoogle is pushing voice agents as a first-class product surface (Gemini Live + Search Live) and backing it with agentic audio evals (ComplexFuncBench Audio, Audio MultiChallenge). Also notable: SynthID audio watermarking baked in.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Voice agents are becoming benchmarkable: what those audio agent benchmarks actually test (and what they miss).”
TurboQuant (extreme compression for KV cache + vector search)
WHY IT ENTERED THE RADARKV cache is the hidden tax of long context. TurboQuant is a crisp, upstream story: “3-bit KV with (near) zero loss” + attention-logit speedups. This has implications for long-context local inference and cheaper serving.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“KV cache is the real cost center: explain TurboQuant in 90 seconds + what it enables (long context on smaller GPUs).”
TurboQuant+ (community implementation + ‘Sparse V’ attention-gated dequant)
WHY IT ENTERED THE RADARThe repo claims a practical speed trick: skip V dequantization where attention weights are negligible (“Sparse V”), reported up to +22.8% decode at 32K on Apple Silicon, with no measurable PPL change in their tests. This is exactly the kind of “small kernel change → big long-context win” that spreads fast.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The next frontier isn’t new models—it’s inference kernels: attention sparsity as compute skipping.”
Voxtral TTS (Mistral’s 4B multilingual TTS + voice cloning)
WHY IT ENTERED THE RADARA lightweight-ish 4B TTS model positioned for real-time voice agents (latency + streaming), plus voice adaptation from ~3 seconds.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“TTS is becoming ‘model + product’: what matters now (TTFA, streaming stability, and controllable emotion).”
Voxtral TTS paper (upstream technical reference)
WHY IT ENTERED THE RADARThe paper spells out the hybrid architecture (AR semantic tokens + flow-matching acoustic tokens) and reports human preference vs ElevenLabs Flash v2.5 in multilingual cloning.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“How modern TTS stacks work (semantic tokens + acoustic tokens) and why ‘3 seconds to clone’ is the new baseline.”
Improving coding agents via “real-time RL” (production feedback → frequent checkpoints)
WHY IT ENTERED THE RADARThis is an upstream training loop story: turn massive production inference into reward signals, ship updated checkpoints as often as every ~5 hours, and deal with reward hacking in the wild.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Your IDE is now a training environment: why coding agents will improve faster than general chat models.”
ARC-AGI-3 (interactive benchmark for agentic intelligence)
WHY IT ENTERED THE RADARBenchmarks are shifting from static Q/A to interactive environments (exploration, long-horizon planning, memory, goal acquisition). If you want a north star for “agents beyond chat,” this is it.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Why ARC-AGI-3 changes the conversation: measuring learning efficiency over time instead of final answers.”
jai (lightweight containment for AI agents running on your machine)
WHY IT ENTERED THE RADARAs “computer use” agents become mainstream, so does the blast-radius problem. jai is a practical middle ground between full VMs/containers and giving agents your real home directory.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“If you let agents run locally, sandboxing becomes the product. Here’s the simplest mental model for containment.”