A Busy Week for AI Agents
This week brought a wave of releases built around AI agents. Liquid AI, Cloudflare, Microsoft, and Meta each shipped tools that let software plan, browse, code, and test on its own. Together, these launches show where AI agents are heading next: smaller, faster, and closer to the device.
Let’s walk through each release and why it matters.
Liquid AI Shrinks the Agentic Model
Liquid AI released LFM2.5-2.6B, a compact model built for AI agents that run entirely on-device. The model carries 2.69 billion parameters. It combines 22 double-gated short convolution blocks with 8 GQA blocks across 30 layers.
That mix matters for speed. According to MarkTechPost, the model handles 131,072 tokens of context. It decodes at 220 tokens per second on an Apple M5 Max chip. It does all this while using under 2.5 GB of memory.
Liquid AI shipped the weights openly in GGUF, MLX, and ONNX formats. Developers can drop the model straight into a laptop or phone. No cloud server required.
This matters because most AI agents today lean on remote servers for every step. A model that plans and calls tools locally cuts latency and cost. It also keeps sensitive data off someone else’s server.
Cloudflare Builds a Browser Just for AI Agents
Cloudflare took a different angle on the same problem. The company introduced Kitesurf, a browser designed specifically for AI agents rather than humans.
Kitesurf runs entirely inside V8 isolates on Cloudflare Workers. It skips Chromium entirely. As MarkTechPost reports, the team built it in just 12 weeks using Rust components like Blitz, Stylo, and Boa JS.
The browser drops features people expect, like tabs and extensions. Instead, it focuses on what AI agents actually need: machine-readable content, scalability, and strict isolation between sessions.
The performance numbers stand out. Kitesurf already passes over 215,000 Web Platform Tests. It uses 3.1 to 3.8 times less CPU than typical browser automation. Memory use drops by 4.7 to 7 times.
Traditional browser automation tools were never built with AI agents in mind. They carry the weight of rendering pages for humans. Cloudflare’s bet is that agents need a leaner, purpose-built tool instead.
Microsoft Automates Unit Testing
Microsoft moved the agent trend into everyday developer work. The company open sourced code-testing-generator, a polyglot unit-test agent. It ships inside the MIT-licensed dotnet/skills repository.
The agent reads a repository before writing anything. It detects the programming language, existing test framework, and coding conventions already in place. Then it figures out the real build and test commands the project uses.
Only after that groundwork does it plan, write, run, and validate its own tests. This step-by-step approach paid off in testing. On Microsoft’s internal benchmark of 152 tasks, the agent completed 140 successfully. Stock GitHub Copilot, running the same underlying model, completed 120.
That gain, roughly 92.1% versus 78.9% task completion, concentrated in specific problem areas. It suggests that context-gathering before coding gives AI agents a real edge over generic prompting.
Meta Brings Agents to the Terminal
Meta Superintelligence Labs entered the coding agent space too. The lab released Muse Code, a terminal-based coding agent, in beta. It runs on a new model called Muse Spark 1.2.
Muse Code plans changes, writes code, and checks results across large codebases. Its async background agents stay active for an entire session. They don’t spawn fresh for every single task.
A local, append-only event log backs the whole system. This makes the agent’s runtime replay-exact. If it crashes, it can restart from exactly where it left off.
Meta trained Muse Spark 1.2 alongside this harness rather than after it. The model focused on long-horizon, repository-scale work from the start. That co-training approach likely explains why the agent handles extended coding sessions so smoothly.
Tuning the Machines Behind the Machines
Alongside these launches, MarkTechPost published a tutorial on Meta’s Ax library for adaptive experimentation. The guide walks through tuning a RandomForest model using Ax’s modern Client API.
The workflow balances predictive accuracy against model size. It defines a mixed search space with integer, float, log-scaled, and categorical parameters. This kind of tuning sits behind many production AI agents, even when it doesn’t make headlines.
What This Week’s Releases Reveal About AI Agents
Look across these five stories and a pattern emerges. AI agents are splitting into specialized tools rather than one giant model doing everything.
Liquid AI focuses on making AI agents small enough to run on a laptop or phone. Cloudflare focuses on giving AI agents a browser built just for them. Microsoft and Meta focus on making AI agents genuinely useful for writing and testing code.
For developers experimenting with these systems, a solid local machine still helps. Running models like LFM2.5-2.6B smoothly benefits from a capable laptop (paid link) with enough RAM to spare.
Takeaways
This week’s launches point toward leaner, more specialized AI agents.
- Liquid AI’s LFM2.5-2.6B runs complex tasks on-device with open weights.
- Cloudflare’s Kitesurf gives AI agents a browser built for machines, not humans.
- Microsoft’s code-testing-generator reads a repo first, then writes better tests.
- Meta’s Muse Code brings crash-safe, long-horizon coding agents to the terminal.
Expect more agent-specific infrastructure like this in the months ahead. The AI agents field is moving past chatbots toward tools built for a different kind of user: other software.
As an Amazon Associate, TechMogo earns from qualifying purchases.
