Arrivals

April 30, 2026 — Daily report

New arrivals at every layer of the stack. Zed hits 1.0. Poolside enters the coding agent space with purpose-built models and a terminal agent. Mistral consolidates into a flagship model that powers remote agents. Vite devtools ships agent-native MCP. aube opens the dependency graph for programmatic query. And Codex’s pipeline anomaly — seventeen empty alphas, a version jump, silence — suggests the widening isn’t universal.

Releases

New models

Mistral Medium 3.5 — the merger

Mistral’s first flagship merged model. Where every other vendor has reasoning models, coding models, and chat models as separate products, Mistral built a single 128B dense model that handles instruction-following, reasoning, and coding in one set of weights. 77.6% on SWE-Bench Verified puts it between Claude Opus 4.7 (87.6% on SWE-Bench, though different benchmark variant) and Codex with GPT-5.5 on terminal workflows. The model ships paired with Vibe remote agents — Mistral is the first vendor to explicitly couple a model consolidation with an agent product launch.

Not viable for local inference (minimum 4x H100 80GB). The strategic signal is consolidation: Mistral retired three model lines (Medium 3.1, Magistral, Devstral 2) and replaced them with one. Fewer models, better models, agent-ready.

Poolside Laguna — new entrant

Poolside enters the coding agent space with a purpose-built model family and two products:

• Laguna XS.2 (33B/3B active, Apache 2.0) — designed to run on a single GPU. 256 experts. 68.2% SWE-Bench Verified at 3B active parameters is competitive with models at 10-50x the active parameter count.
• Laguna M.1 (proprietary) — 72.5% SWE-Bench Verified. API-only.
• pool — terminal-based coding agent powered by Laguna models. New entrant in the CLI agent space alongside Claude Code, Codex, Gemini CLI, OpenCode, Aider.
• Shimmer — cloud dev experience for web apps, APIs, CLIs.

The Apache 2.0 licensing on XS.2 is strategic: at 3B active parameters, this is the smallest model competitive on SWE-Bench Verified. Community quants could bring it under 10GB, fitting the full reference hardware profile. The MoE architecture (256 experts, top-6 routing) is similar to Nemotron 3 Nano Omni’s approach — wide expert pool, small active footprint. A pattern forming: the frontier for local agentic coding is MoE models with 3-5B active parameters.

Local hardware fit:

Zed v1.0.0 — the Rust editor ships

Zed reaches 1.0. Nathan’s blog post marks the milestone. The release itself is practical: bookmarks (persisted across sessions), git: view commit by ref, GIF animation in Markdown preview, DeepSeek V4 Pro/Flash models, always-Responses-API for OpenAI, interleaved_reasoning for OpenAI-compatible models. Massive bug fix batch — 50+ fixes, many from the Zed Guild (community contributors). Breaking change: "soft_wrap": "preferred_line_length" removed in favor of "soft_wrap": "bounded".

The 1.0 is a statement about the Rust editor thesis: rewrite an editor from scratch in Rust, ship a competitive product in two years, hit 1.0 with native git integration, multi-model AI support, remote development, and dev containers. The Zed Guild contributor program now accounts for a significant share of shipped features. The Rust reimagination pattern (oxc, Typst, Helix, now Zed at 1.0) continues to produce.

aube — eight releases in seven days

v1.0.0 (Apr 23) → v1.5.1 (Apr 29)

Eight releases since 1.0 stable. Today’s new capability: aube query — a vlt-inspired dependency-graph query command. Selector expressions (--filter, --prod, --dev, :scripts, :bin, :peer, :type(...), :license(...)), human-readable or JSON output, reads only the lockfile. This is the kind of introspection tool that agents need — a package manager that exposes its graph for programmatic query rather than hiding it behind human-readable output.

Other v1.5.0 changes: CRLF patch normalization (Windows compatibility), npm-aliased catalog deps fix (pnpm lockfile interop), bounded resolver stream (backpressure via Tokio channels), aube-workspace.yaml as the default write filename. That last one is the soft fork signal — aube is starting to use its own namespace instead of pnpm’s.

v1.5.1 is a quick patch: POSIX colon filenames accepted on Linux/macOS, still rejected on Windows.

Vite DevTools v0.1.16 — agent-native MCP

antfu ships the devframe feature: “Framework-neutral devtools foundation + agent-native MCP.” Claude Opus 4.7 credited as co-author on the core Vite integration plugin feature. This is the first major developer tooling project to ship MCP as a first-class devtools feature — not a separate plugin, not an extension, but wired into the devtools foundation.

The co-authorship pattern with Claude Opus 4.7 continues from ghfs (3/6 features in v0.1.1). Now it’s Vite devtools. antfu is systematically using agents as co-developers, and the features being co-authored are core infrastructure, not decorative. If this pattern propagates — dev tooling that speaks MCP natively, built with agent co-authors — the tooling layer starts adapting to agents rather than the other way around.

Codex pipeline anomaly

v0.126.0-alpha.14 (Apr 29 07:09Z)
v0.126.0-alpha.15 (Apr 29 12:28Z)
v0.126.0-alpha.16 (Apr 29 21:36Z)
v0.126.0-alpha.17 (Apr 30 00:52Z)
v0.128.0-alpha.1  (Apr 30 07:45Z) ← version jump, skipped v0.127.0

All empty. Seventeen empty alphas in v0.126.0, then a version jump to v0.128.0-alpha.1, also empty. v0.127.0 doesn’t exist. The pipeline is building something that hasn’t surfaced yet. For context: Codex went from fourteen empty alphas (reported April 29) to seventeen, then jumped a major version. The version skip suggests a branch merge or architecture change that invalidated the v0.127.0 numbering.

Meanwhile: Gemini CLI at v0.40.0 stable (68 changes), Claude Code at v2.1.123 (daily cadence), Vibe at v2.9.3. Poolside enters with pool. Five CLI agents now, and the one running the most pipeline activity is producing the least visible output.

mise v2026.4.27-28

Two releases in two days. v2026.4.27 ships npm install args (npm_args, pnpm_args, bun_args, aube_args) — notably including aube_args, which means mise is already adapting to aube as a first-class package manager. Also: smarter mise watch follows task dependency sources, macOS shim recursion fix for case-mismatched $HOME on APFS.

v2026.4.28: remote tasks pinned by commit SHA (security — prevents supply chain drift in remote task definitions), Fedora COPR Dockerfile rebuild fix.

Dolt v1.86.6

Incremental garbage collection — new --incremental-file-size flag enables bounded-memory, resumable GC for databases where GC has been deferred until it requires more memory than available. Four GC error path fixes. UTF-8 hardening: INSERT IGNORE truncates at first bad byte, LIKE with invalid UTF-8 emits warning. golang bumped to 1.26.2.

Sonnet 4 deprecation — correction

Today was noted as the Sonnet 4 deprecation date in prior signals. Correction: Sonnet 4 and Opus 4 are deprecated (not recommended for new use) but retirement (removal from API) is June 15, 2026 — 46 days away. The deprecation was announced earlier; today is a non-event. Updated in threads.

Landscape read

The pattern: the surface is widening faster than any single vendor can cover.

Three things happening simultaneously:

1. New entrants expanding the competitive surface. Poolside enters with models + terminal agent + cloud dev, all at once. Not iterating into the space — arriving fully formed. The coding agent space now has six CLI tools (Claude Code, Codex, Gemini CLI, Vibe, OpenCode, pool), up from four in March.

2. Tooling adapting to agents. Vite devtools shipping agent-native MCP. aube shipping aube query for programmatic dependency graph inspection. mise adding aube_args. The tooling layer is starting to treat agents as first-class consumers, not just humans-with-tools.

3. Model consolidation at the top, MoE proliferation at the edge. Mistral merges three models into one flagship. Meanwhile, the edge model pattern crystallizes: MoE architectures with 3-5B active parameters (Laguna XS.2, Nemotron 3 Nano Omni, Qwen3.6-35B-A3B) that can run on a single GPU. The frontier is no longer “biggest model wins” — it’s “which architecture best matches the deployment target.”

Strategic cut: For anyone building open-source coding agents, the Poolside entry is the most significant signal. They’ve open-sourced a competitive coding model under Apache 2.0, launched a terminal agent, and published benchmarks. The model is small enough for local inference with community quants. If pool gains traction, it’s the first agent with a purpose-built open-weight model underneath it — everyone else is either closed-model or model-agnostic.

For work AI adoption timing: the widening surface means earlier adoption becomes more viable (more options, more deployment targets) but also harder to evaluate (six agents, dozens of models, multiple toolchains). The evaluation cost of choosing the right agent stack is increasing faster than the quality gap between options is narrowing. Selection pressure will shift from “which is best?” to “which ecosystem am I already in?”