Which developer tools for AI agents stood out in the first half of March 2026?

In the first half of March 2026, OpenAI explained how its Responses API gives AI agents a hosted computer with a shell and files (11 March). The MASEval paper (9 March) found that the agent framework can matter as much as the model. On GitHub, Gemini CLI passed 98,000 stars and Everything Claude Code grew from about 70,000 to more than 86,000 stars in nine days. Anthropic's programmatic tool calling, in which Claude calls tools from code, is older: it dates from November 2025.

How does OpenAI give agents a computer in the Responses API?

In an engineering post of 11 March 2026, OpenAI described how the Responses API runs agents in a computer environment. The model proposes shell commands and the platform runs them in a hosted container: an isolated workspace with a filesystem for inputs and outputs, optional structured storage such as SQLite, and restricted network access. Models from GPT-5.2 onward are trained to use this shell tool. The API can run several commands in parallel, and it enforces an output limit that the model sets per command.

Outbound traffic passes through a proxy that enforces allowlists. Secrets such as API keys are only added on the way out, per domain, so the model and the container see placeholders. For long tasks the API offers compaction: the model condenses the earlier conversation into a compact item, so the work can continue when the context window fills up. Skills, folders with a SKILL.md file and scripts, are copied into the container, where the model can find and run them. OpenAI's point is that developers no longer have to build and maintain such an execution environment themselves.

What is Anthropic's programmatic tool calling?

With programmatic tool calling, Claude writes Python code that calls your tools and runs it in Anthropic's sandboxed code execution container. Normally every tool call is a round trip through the model: Claude asks for a call, you return the result, and the result lands in its context. Here the code makes the calls, filters the results and passes on only its final output. When the code calls one of your tools, execution pauses and the API returns an ordinary tool call. Your result goes back into the code instead of into Claude's context. You switch it on per tool with the allowed_callers field, and ordinary function calling keeps working.

The feature was not new in March 2026. Anthropic introduced it in beta on 24 November 2025, and since 17 February 2026 it works without a beta header. In Anthropic's internal tests on complex research tasks, average token use fell from 43,588 to 27,297, a reduction of 37%.

What does Gemini CLI offer?

Gemini CLI is Google's open-source AI agent for the terminal, under the Apache 2.0 license. In March 2026 its README listed Gemini 3 models with a context window of 1 million tokens, built-in tools for Google Search (to ground answers in search results), file operations, shell commands and fetching web pages, and support for MCP servers for your own integrations. With a personal Google account, the free tier allowed 60 requests per minute and 1,000 requests per day. On 17 March 2026 the repository had 98,090 stars and 12,349 forks, according to a Wayback Machine snapshot.

What does MASEval measure?

MASEval (Emde et al., 9 March 2026) starts from a complaint: most agent benchmarks fix the setup around the model and compare only models. MASEval evaluates the whole multi-agent system, including its topology, orchestration logic and error handling. The authors ran every combination of three frameworks (smolagents, LangGraph and LlamaIndex), three models (Gemini 3 Flash, GPT-5 mini and Claude Haiku 4.5) and three benchmarks: 27 configurations. Their conclusion: among models of comparable capability, the choice of framework matters as much as the choice of model. The library works with any framework and is available on GitHub under the MIT license.

What is Everything Claude Code?

Everything Claude Code (ECC) is a community project by developer Affaan Mustafa, who according to the README won an Anthropic hackathon. It calls itself a performance optimization system for agent harnesses: programs such as Claude Code that run a coding agent. It bundles skills, subagents, slash commands, hooks that save and load context between sessions, and a security scan called AgentShield. In March 2026 the README listed 25 agents, 108 skills and 57 commands for Claude Code; the OpenCode version had fewer, for example 37 skills. Version 1.8.0 (5 March 2026) improved the reliability of hooks and made behaviour more consistent across Claude Code, Cursor, OpenCode and Codex. According to Wayback Machine snapshots, the repository went from 70,041 stars on 10 March to 86,438 on 19 March 2026.