Poolside's Laguna S 2.1 Beats DeepSeek With One-Sixth the Active Parameters

The chat box was never the right interface for a coding agent doing real work. Poolside just shipped Poolside Desktop Assistant, a macOS app that treats agent supervision as a first-class activity. It comes paired with a model that beat DeepSeek V4 Pro Max on agentic coding benchmarks while activating a fraction of the parameters.

Why the chat column breaks down

When a coding agent is reading through a repository, running tests, fixing what it broke, and grinding away for hours, a single scrolling chat column stops being useful. The real unit of work is a session tied to a repository, a branch, its tools, its permissions, a runtime, and everything that's happened so far. And there's rarely just one session running. You might have one agent implementing a feature, another investigating a bug, and a third reviewing the first agent's output. Your job becomes supervision, not prompting.

Poolside built this tool for their own team first. For the past year, the team did its daily work inside tools they built for themselves: pool, their agent harness and CLI, and the interfaces around it. Now they're making those interfaces available to everyone.

One workspace, any agent

The headline feature is vendor-agnosticism. Poolside Desktop Assistant runs multiple coding agents across projects and repositories, with Poolside Assistant extensions bringing the same experience into VS Code and Visual Studio.

The app is built on the Agent Client Protocol (ACP), which works like LSP but for agents instead of language servers. Created by Zed Industries and released in August 2025, it uses JSON-RPC 2.0 over stdin/stdout to create a universal interface between agents and editors. Before ACP, every agent-editor combination required a custom integration. ACP standardizes that interface so any compatible agent works in any compatible editor, no custom glue code required. It's now adopted by JetBrains, Google, GitHub, and 25+ agents.

In practice, you can bring pool, Claude Code, Codex, or Gemini into one workspace and customize the layout around how you work. No Poolside account required to use the app with third-party agents.

Features worth knowing

• Parallel agents with Git worktree isolation. Run multiple sessions in parallel, each isolated in its own Git worktree, so agents can work concurrently without fighting over the same files. You decide which changes are worth merging.
• Agent handoffs with context intact. Conversations are not welded to the agent that started them. Hand off a session along with its full context: have one agent write the plan and another implement it, ask Claude Code to review what pool built, or hit a rate limit and carry on with a different harness.
• A layout built for supervision. Supervising agents means moving between the conversation, terminal, file tree, current diff, and repository state. Poolside Assistant treats those as parts of one workspace. Arrange tabs, splits, a sidebar, and a bottom panel, then save your preferred layout as default.
• Session history import. If you already use Claude Code or Codex, import your existing sessions from Settings → Archive and pick up where you left off.
• Built-in connectors. Poolside Assistant ships with connectors for Exa, Parallel web search, GitHub, Linear, Notion, Coda, Granola, Sentry, New Relic, Vercel, Supabase, Hugging Face, and more.
• Fully offline mode. Local model support runs through MLX. Download Laguna XS 2.1 for a completely offline setup, or search for other MLX-compatible models directly from Hugging Face. The app estimates how well a model fits your Mac's unified memory before you download it.

The model underneath: Laguna S 2.1

Out of the box, the app pairs with Laguna S 2.1, Poolside's flagship open-weight coding model released days before the app. Laguna S 2.1 is a 118B Mixture-of-Experts (MoE) model with 8B active parameters, a 1M context window, and a thinking mode, scoring 70.2% on Terminal-Bench 2.1.

Terminal-Bench 2.1 measures an agent's ability to complete real, long-horizon terminal tasks autonomously, making it closer to production use than static code-completion benchmarks. That 70.2% places Laguna S 2.1 ahead of DeepSeek-V4-Pro-Max (a 1.6-trillion-parameter model that scored 64.0), Thinking Machines' 975-billion-parameter Inkling (63.8), and Nvidia's 550-billion-parameter Nemotron 3 Ultra (56.4). On DeepSWE, the efficiency gap is even wider: Laguna S 2.1 scores 40.4% against DeepSeek-V4-Pro-Max's 9.0%, with roughly one-sixth the active parameters.

MoE architecture is why this is possible. The model contains many specialized sub-networks called experts, but only activates a small subset for each token. Total parameter count stays large; inference stays fast and cheap. Pre-training began May 22, 2026, on 4,096 Nvidia H200 GPUs, and S 2.1 is the first Poolside model trained with reinforcement learning in FP8 precision. The full run, from training start to launch, took under nine weeks.

There are known rough edges. In unfamiliar environments with slightly different tool schemas, the model can drift from the required output format. Closed frontier models such as Claude Fable 5 and Kimi K3 still lead on several benchmarks.

What this means for the field

Poolside Desktop Assistant is a bet that the interface layer for coding agents should be decoupled from the model layer, the same way LSP decoupled language intelligence from IDEs. ACP is the open standard making that possible. If it gains traction, the competitive moat for any single agent provider narrows: you keep users by being the best agent, not by locking them into your interface.

The parallel-agent-with-worktree-isolation pattern is also worth watching. Running multiple agents concurrently on isolated branches, then reviewing and merging their outputs, is a workflow that lacked first-class tooling until now. Because the interface, agent harness, and model stay separate and composable, the pattern works regardless of which agents you trust for which tasks.

Getting started

Poolside Desktop Assistant is available today for macOS, VS Code, and Visual Studio. You can also install pool to work from the terminal, or connect Poolside agents to Zed and JetBrains through ACP. Laguna S 2.1 and Laguna XS 2.1 are free in Preview through the Poolside get-started page. Laguna S 2.1 is also available free on OpenRouter at 256K context, with full weights open under the OpenMDW-1.1 license on Hugging Face for self-hosting.

The app requires no Poolside subscription when used with third-party agents. Bring your own Claude Code or Codex subscription and the app is just the workspace. Poolside is collecting feedback through a dedicated Discord channel for bugs and feature requests.