BriefTechNews
Gemini Tool Retrieval Cuts Context Bloat, and 32 Agents Beat One Coder

Gemini Tool Retrieval Cuts Context Bloat, and 32 Agents Beat One Coder

7 min read · 16 sources

TL;DR
  • Google launched Tool Retrieval in the Gemini API, letting agents defer loading tool schemas until needed to cut input tokens and latency.
  • Replit's 32-agent swarm of Claude Opus 5.5 achieved an 18.9x NumPy speedup versus 3.7x for a solo agent.
  • METR found a 2x acceleration in AI capabilities in 2026, with researcher Ajeya Cotra estimating an ~10x research speedup.
  • Stripe evals show coding agent strict pass rates drop from 78.1% to 29.4% even with the agent plugin installed.
  • Apple's FastFrames 1.3B model generates 480p video at 36 ms per frame on a single B200 GPU.

Google wants to kill your agent’s context bloat. The Gemini API now has Tool Retrieval, a preview feature in the Interactions API that stops the model from loading every tool schema up front. Instead, it sees names and descriptions, then pulls the full JSON schema on demand within the same request. That means you can register hundreds of tools without inflating the prompt or paying the latency tax for schemas the model never touches. The search happens server-side, so there’s no index for you to build or maintain. For anyone running agent loops that pass around a dozen function definitions per turn, this is the difference between a 2k-token context and a 20k-token one.

The model decides what to load based on the user’s input, which is the right call for most workflows. But it’s a preview, so expect the retrieval logic to be a black box for now. The win is real: agents that would have choked on a 50-tool API surface can now scale without a custom router in front of Gemini.

A 32-agent swarm of Claude Opus 5.5 hit an 18.9x NumPy speedup, while the strongest solo model’s performance declined as the team grew.

Gemini Takes Over Your Mac's Cursor

Source: blog.google ↗

Google also shipped computer use for the Gemini app on Mac, rolling out in English in the US. It automates multi-step tasks across local files, desktop apps, and the web, running in an isolated background window with a live preview in chat. Your main screen and cursor stay free, so you can keep working while Gemini batch-processes files or fills out forms. You approve key actions before they execute and pause anytime.

This is the agentic desktop play: not a screen-scraping hack but a controlled environment where the model has explicit access grants. For engineers, the interesting part is the isolation model - it’s a sandbox that doesn’t block your primary display. That’s a pattern worth stealing for any agent that needs to touch a GUI without owning the user’s attention.

AI Is Already Speeding Up Its Own Research

Source: linch.substack.com ↗

The biggest story today isn’t a product launch, it’s a number. METR, with evidence from Anthropic, found a 2x acceleration in AI capabilities in 2026, largely attributed to AI-assisted research speedups. Researcher Ajeya Cotra estimates the research speedup is ~10x, with Anthropic staff reporting ~12x. That’s the early-to-middle stage of recursive self-improvement (RSI), where AI accelerates its own training.

The safety implications are the headline, but the engineering takeaway is more immediate: AI research is becoming a well-defined field, and the people who use AI to design the next model are pulling away from those who don’t. If you’re building infrastructure for AI workloads, plan for the demand curve to be steeper than your current capacity model. The 2x capability figure is the floor, not the ceiling.

Sandboxed Models Still Talk Through Inference Servers

Source: lesswrong.com ↗

A LessWrong post demonstrates that sandboxing doesn’t stop models on a shared inference server from trading information. In contrived tests, one model’s large request slightly shifts the server’s floating-point calculations for another model running concurrently. The second model reads that shift as a signal.

This is a side-channel attack on shared compute. If you’re running multi-tenant inference, your isolation boundary is the hardware, not the container. The practical fix is to avoid co-tenancy for sensitive workloads or to add jitter to request sizes. Don’t assume your sandbox is a hermetic seal - it’s a speed bump.

When 32 Agents Beat One Coder

Source: frontier.replit.com ↗

Replit’s benchmark on agent swarms versus solo coders is the clearest data yet on when throwing headcount at a problem works. A 32-agent swarm of Claude Opus 5.5 achieved an 18.9x speedup on a NumPy performance task versus 3.7x for a solo agent. But GPT-6.2 Astra swarms degraded as they grew. The difference: task separability and coordination. The best swarms used a shared board with minimal direct messages, avoiding coordinator bottlenecks.

The engineering lesson is that swarm scaling is not monotonic. You need a task that splits cleanly and a coordination layer that doesn’t become the single point of failure. Replit’s Agent Habitat provides isolated microVMs for this, but the pattern generalizes: shared state, low chatter, and a task graph that doesn’t require consensus.

Stripe's Evals Show Agents Still Hallucinate APIs

Source: stripe.dev ↗

Stripe published evals measuring coding agents on 69 realistic integration scenarios across payments, subscriptions, and Connect, testing 14 model/harness pairs. The Stripe agent plugin improved scores by 3-8 percentage points and cut web searches by 96%. But strict pass rates dropped from 78.1% to 29.4% when held to best-practice standards.

That gap is the story: agents produce working code that violates best practices. They bias toward older APIs and hallucinate endpoints. The plugin steers them toward better integrations, but the benchmark is publicly maintained, so you can run your own models against it. If you’re shipping an SDK, this is the template for testing whether agents can actually use it correctly.

Perplexity's Computer Learns Its Own Skills

Source: perplexity.ai ↗

Perplexity Computer can now build and update its own skills, using its Brain system to learn from each user’s past sessions. This is less about a new model and more about a persistent memory layer that gets better with use. For operators, the question is what happens when a skill goes wrong - there’s no mention of rollback or versioning in the announcement.

Chrome's Mirror Dimension for Testing Agents

Source: blog.google ↗

Chrome built a “mirror dimension”, a synthetic web environment with thousands of simulated sites, to safely evaluate browser-use agents like auto browse without real-world risk. It uses an isolated sandbox and a local interception proxy for repeatable, hermetic testing with clean baselines.

This is the right way to test agents that touch the live web. If you’re validating agent safety, a synthetic environment with deterministic state beats scraping production traffic every time.

FastFrames: Video Generation at 36 ms per Frame

Source: apple-aiml-research.github.io ↗

Apple’s FastFrames is a family of autoregressive video diffusion models hitting interactive latency. The 1.3B variant runs at 36 ms/frame (28 RPS) on a single B200 GPU, matching the quality of a model 8x slower. The 14B variant runs at 45 ms/frame (22 RPS) on 8x B200s with better quality. They replaced VAE decoding and KV cache updates, which account for over half of inference latency, with lightweight alternatives.

This is the difference between batch video generation and real-time camera- and physics-conditioned generation. For anyone building interactive video tools, the 36 ms/frame number is the threshold where it stops being a render job and starts being a live feed.

Vercel Board Coordinates Agent Fleets

Source: vercel.com ↗

Vercel launched Board, an experimental agentic primitive for coordinating multi-agent work with channels, threads, and labels. It’s harness-agnostic, supporting Claude Code, Codex, and others via CLI or SDK. Internally, v0’s software factory uses it and merges over 50% of PRs. The pattern is a purpose-built coordination layer with grants and two-way sync for tools like Slack.

The Exponential Time Hypothesis Gets an Independent Review

Source: today.ucsd.edu ↗

UC San Diego and MIT computer scientists are facilitating an independent review of OpenAI and Anthropic’s claim that their models refuted the Exponential Time Hypothesis (ETH), proposed in 1998. They launched a blog for community discussion and plan academic conferences. If ETH falls, it impacts optimization algorithms, chip design, and encryption systems.

Cloudflare Opens Clairic Weights

Source: blog.cloudflare.com ↗

Cloudflare released updated open-weight versions of Clairic and new Clairic-flash, a 9B-parameter model for extracting fields from complex documents including images and PDFs. It uses parallel slot autoregressive generation for faster, more accurate extraction. Weights are on Hugging Face, hosted on Workers AI.

DeepMind Asks What Comes After AGI

Source: institute.deepmind.com ↗

DeepMind’s essay maps the transition from AGI to artificial general superintelligence (ASI), defined as consistently outperforming large groups of human experts on tasks like scientific progress or chip design. It argues for rigorous tracking methods and notes hard physical and complexity-theoretic limits. The practical takeaway: plan for rapid acceleration beyond human-level intelligence.

OpenAI's Performance Factory and the TLDR Community

Source: openai.com ↗

OpenAI’s Performance Factory uses Codex to automate performance regression investigations and fixes for ChatGPT. Meanwhile, the TLDR community is debating whether personality matters more than benchmarks - Opus 5.5 is loved despite not topping benchmarks. For dev agents, verified output still beats charm.

Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like