VT Code debuts as Rust-native terminal coding agent

TLDR: New Rust TUI coding agent, OpenAI GPT-5.5, and DeepSeek V4 are all quietly ratcheting up what “agentic” actually means in production as of 2026-04-26.

VT Code ships Rust terminal coding agent with MCP support

VT Code is a new Rust based terminal user interface (TUI) coding agent that speaks to Anthropic, OpenAI, Google Gemini, and open source models, with local inference via LM Studio and experimental Ollama support. It uses semantic context via ast-grep for structured code search and ripgrep for fast text search, so the agent can operate on your repo with more structure than a raw grep loop.

For engineers who live in a terminal and want an agentic workflow without a heavy IDE, this is worth a look. VT Code is built on Ratatui, and the architecture plus agent loop are documented in the README, which matters if you plan to extend or audit it. Model Context Protocol (MCP) and Agent Client Protocol (ACP) support mean you can plug it into a broader tool ecosystem instead of yet another siloed wrapper.

Still early, but the choice of Rust, explicit semantic tools, and protocol support is a strong signal of where serious agent shells are going. If you are exploring repo-native agents or want to standardize on MCP compatible tools, this is a concrete reference implementation. As of 2026-04-26 there are no formal benchmarks, so evaluate it on your own codebases.

Read more →

---

OpenAI GPT-5.5 posts stronger agentic coding benchmarks

OpenAI GPT-5.5 is reported to score 82.7% on Terminal-Bench 2.0 and 73.1% on Expert-SWE, signaling measurable progress on agentic tasks such as coding and terminal workflows. The model reportedly cuts token usage while holding latency roughly constant, which directly improves cost profiles for high volume production agents.

For teams running code assistants, CLI agents, or workflow copilots, these benchmark jumps matter more than raw chat scores. Better tool use, reduced hallucinations, and improved steerability make it easier to trust the model for knowledge intensive domains such as scientific research or legal analysis. The catch: these are still vendor supplied or vendor curated benchmarks as of 2026-04-26, so you should validate against your own evals and guardrails.

If the pricing lands close to current GPT-4.5 tiers, expect plenty of silent backend upgrades across products with “now smarter” release notes. The main question is how much of the claimed hallucination reduction holds up under messy, real enterprise data.

Read more →

---

DeepSeek V4 preview targets cheap long-context agent workloads

DeepSeek V4 Flash and V4 Pro Max are rolling out in preview with both API access and an open source release, which means you can start integrating them into long context workloads immediately. Early benchmarks suggest V4 Flash could become a default choice for cost sensitive production, while V4 Pro Max goes after high end reasoning and large scale analysis.

For anyone building retrieval-augmented generation (RAG) over big codebases or multi step reasoning agents, long context plus decent reasoning at lower cost is the real draw. The open source angle and support for Chinese hardware stacks, including Huawei Ascend as covered elsewhere, also give non U.S. ecosystems a more competitive option. As of 2026-04-26, community evals are still sparse and the models are no longer benchmark leaders, so expect some tradeoffs against U.S. frontier models.

If DeepSeek keeps iterating with public feedback, V4 may become the “good enough and cheap” default that powers a lot of behind the scenes automation where vendor lock in or export controls are concerns.

Read more →

Quick Hits

• Stanford Medicine-Led Study Shows AI Enhances Physician Medical Decision-Making This Stanford Medicine study reports that clinicians using AI support made better treatment decisions in complex cases, which is a useful datapoint if you are designing clinical agents or selling into healthcare as of 2026-04-26.

• Show HN: Agent MCP Studio – build multi-agent MCP systems in a browser tab A browser only studio for authoring tools, orchestrating multi agent MCP systems, retrieval augmented generation, and code execution, all inside a single static HTML file via WebAssembly and Pyodide, useful for safe experimentation without a backend.

• Show HN: Nimbus – Browser with Claude Code UX Nimbus is a desktop browser that bakes in an AI agent with a Claude Code inspired chat bar UX, treating the URL bar as mostly redundant and giving you a reference for agent native browsing interfaces.

• The people do not yearn for automation Simon Willison highlights Nilay Patel’s essay on why AI remains unpopular with the public despite skyrocketing usage, worth reading if you are building agents that replace or reshape human workflows.

• Millisecond Converter A tiny utility from Simon Willison that converts millisecond timings into seconds and minutes so you can reason about LLM latency without doing mental math.

• llm-openai-via-codex 0.1a0 A plugin that lets the llm CLI talk to OpenAI models using existing Codex subscription credentials, handy if your organization still has legacy Codex access wired into tooling.