🔁 Agentic Edition: Agent runs you can rewind, MCP tools over zero-trust P2P, and a tool caller that fits in 28MB of RAM

★ Read this first

First issue of the agents list. One rule for what gets in: it has to change what an agent can do, not what a model can score.

Read the six together and the same shape appears. Nobody shipped a smarter agent this month. They shipped the things around it — a network, a runtime, a browser, a body, a training environment, an undo button. The model has been the easy part for a while. This is the month the rest of it arrived.

96.8%

Kernels beat compiler

14MB

Whole tool caller

4.7–7×

Less RAM than Chromium

01 Google · Agent networking

Meet SAM (Sovereign Agent Mesh): A Zero-Config, Zero-Trust P2P Network for AI Agents

Two agents on two networks share a tool by publishing an endpoint to the internet. That is the whole security problem, and it is why most internal tools never get shared at all. SAM is a P2P overlay instead. Nodes find each other, survive NAT, and nothing internal faces outward. It is not Segment Anything, and the repo says plainly it is not an officially supported Google product.

The identity path is the interesting part. The control plane verifies an OIDC JWT and seals the claims into a Biscuit token as Datalog facts, so a node authorizes a call offline without calling home. Default-deny is absolute — even the discovery catalog needs an explicit grant. Apache-2.0, three Go binaries, Helm chart included. The public mesh is still a beta testnet, so production means self-hosting your own control plane.

Read on Marktechpost →  ◆ GitHub  ◆ Docs  ◆ Kubernetes guide

02 DeepSeek · Agent runtime

DeepSeek AI Releases DeepSeek Harness in Developer Preview: An MIT-Licensed Agent Harness Where Everything is a Plugin

Every harness hard-codes something. The agent loop, the tool registry, the session store — you extend at whatever hooks the authors thought to leave you. DeepSeek states the opposite in the first lines of its README: everything is a plugin. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI all sit behind Cordis plugin boundaries, swappable in configuration without touching Harness source.

Four modes load different default plugin sets: Standard, Code, Minimal (two tools — a persistent bash and str_replace_editor, for benchmarking in a bare environment) and Creator. Every context injection lands in an append-only session log, not just tool calls, and resume, fork, search and replay all read the same event stream. MIT, ships as dsh, v0.1 developer preview.

Read on Marktechpost →  ◆ GitHub  ◆ Product page  ◆ Docs  ◆ Cordis

03 Cloudflare · Agent browser

Cloudflare Introduces Kitesurf: An Agent-First Web Browser That Runs Entirely in V8 Isolates on Cloudflare Workers

Chromium was built for a person looking at a screen. An agent does not need tabs, extensions, or pixel-perfect 60-fps rendering, but it pays for all three — which is what makes one-browser-per-agent expensive. Kitesurf drops the human parts. It runs entirely in V8 isolates on Workers with no Chromium underneath, using Blitz and Stylo for HTML and CSS, and it already passes 215,000+ Web Platform Tests.

On Cloudflare’s 14-URL corpus: 380ms CPU per screenshot against 1,173ms for Chromium, and 271.0 MiB of memory down to 57.8 MiB. HTML extraction drops from 273.7 MiB to 39.4 MiB. Chromium is still faster on wall time by roughly 1.7–1.8×, mostly in rasterization — the trade favors bursty agent work because memory and CPU are what you get billed for. Existing Puppeteer, Playwright and MCP clients switch by adding browser=kitesurf. No video, WebGL, or persistent authenticated sessions yet.

Read on Marktechpost →  ◆ Announcement  ◆ Docs  ◆ Live playground

04 Cactus Compute · On-device tools

Meet Needle 2: An Open 45M-Parameter Tool-Calling Model That Ships as a 14MB Binary and Runs a Full Session in 28MB of RAM

The premise is stated plainly by the team and it is the whole argument: mapping a messy sentence onto a typed function signature needs no world knowledge and no open-ended prose. So 45M parameters are enough, and the target is hardware with no GPU and no NPU. The model is sealed inside Cactus’s own C++ engine — no runtime to install, no download at inference time.

It leads both Seal-Tools splits on ordered strict exact match — 32.6 in-domain and 28.7 out-of-domain, against 26.9 and 17.0 for LFM2.5 230M — and trails on BFCL v4 overall at 42.6 versus 61.7 for Apple FM. Cactus flags the asymmetries itself: baselines run f16, and task specialization favors Needle. A 256-token sliding window with pinned tool sinks keeps RAM near 28MB no matter how long the conversation runs.

Read on Marktechpost →  ◆ GitHub  ◆ Hugging Face  ◆ Paper  ◆ Product page

05 ByteDance Seed & Tsinghua AIR · Agentic RL

ByteDance Seed and Tsinghua AIR Introduces CUDA Agent: A Large-Scale Agentic RL System for CUDA Kernel Generation

Frontier models already write correct CUDA. They write slow CUDA. On KernelBench the base model Seed1.6 passes 74.0% of tasks but beats torch.compile on only 27.2%, at 0.69× geomean — on average slower than the compiler. The fix was not a better prompt. It was putting the model inside a real CUDA environment with a profiler, correctness checks and a permission-locked sandbox, then training with PPO for 150 steps.

Table 1 reports 98.8% pass and 96.8% faster-than-torch.compile across the 250 tasks, at 2.11× geomean. The ablation is the sentence to keep: remove the agent loop and the faster rate collapses from 96.8% to 14.1%. Weights are closed — the trained agent sits on proprietary Seed1.6 — but the 6,000-sample dataset, the SKILL.md spec and the reward recipe are public.

Read on Marktechpost →  ◆ Paper  ◆ Project page  ◆ Dataset

One more thing…

Northeastern & Stanford · Agent substrate

Meet Shepherd: An Open-Source Python Substrate That Lets Meta-Agents Fork, Replay, and Revert Any Agent Run

Your agent is at step ten. It misreads a traceback and rewrites a file that was already correct. You have two options and you hate both. Patch forward, and the context grows and the token bill grows with it. Restart, and you re-pay every model and tool call while reproducing nothing exactly, because the run was never deterministic.

What you actually want is step eight. No runtime could give it to you, because Git versions files — not a live process, not a dev server, not a warm prompt cache. Shepherd records every agent-environment interaction as a typed event in a Git-like trace, and each commit covers the agent process and the filesystem together, copy-on-write. A rewind restores live state, not just files.

The research team reports forks 5× faster than Docker and over 95% prompt-cache reuse on replay. Once a run is forkable a meta-agent can sit on top of it: a live supervisor raised CooperBench pair-coding pass rates from 28.8% to 54.7%, and forking rollouts at selected turns moved TerminalBench-2 from 34.2% to 39.4%. MIT-licensed, pip install shepherd-ai, Python 3.11+, early alpha.

Read on Marktechpost →  ◆ Paper  ◆ PyPI  ◆ Project page

Sponsor AI Dev Signals

Reach the people shipping agents, not the people posting about them.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

The agents edition ships with four placements per issue. This one is open.