How agents-exe compares
On Sat, 03 Oct 2026, by @lucasdicioccio, 644 words, 0 code snippets, 14 links, 0images.
How agents-exe compares
This page puts agents-exe next to five agent frameworks people are likely to be choosing between: LangGraph, AutoGen, the OpenAI Agents SDK, the Claude Agent SDK and CrewAI.
Read this first. It is a draft. What it says about agents-exe was checked against the code or by running it. What it says about the other five comes only from their own documentation, read on 2026-10-03 and linked under Sources; none of them was installed or run for this page. There are no benchmarks here, and no claim that one is faster, cheaper or better than another. A cell that says “not checked” means exactly that, not “missing”. Which frameworks belong on this page, and on which axes, is still an open choice; corrections are welcome as issues.
What agents-exe is
- A command-line program, configured with files. An agent is a JSON file: an endpoint, a model, a system prompt, tools, helper agents. There is no Python or TypeScript API; the library underneath is Haskell.
- Tools are programs. Any executable that answers
describeandrunis a tool, whatever language it is written in. MCP servers, OpenAPI documents and a few builtin toolboxes (SQLite, system information, Lua) are tools too. - Helper agents are tools as well. A root agent calls a helper the way it calls a tool, and helpers can have helpers.
- One agent file, several ways to run it: a one-shot command, a terminal UI, a stdio MCP server, or an HTTP server with a chat page.
- Sessions are stored. As JSON files for the command line, in SQLite or PostgreSQL for the server; a session can stop on a tool call that waits for an outside answer and be resumed later.
- Self-hosted. It talks to the model endpoint you give it; there is no agents-exe service. Apache-2.0.
The demo shows several of these at once.
What it is not
- Not a library for a Python or TypeScript application: if you want to build agents as code inside such an application, the five frameworks below are made for that and agents-exe is not.
- Not tied to one model vendor, and not a client for every vendor’s native API either: models are reached through OpenAI-style chat-completions endpoints (the README lists OpenAI, Mistral, Moonshot and Ollama as working).
- Not a hosted product: you run the binary and keep the database.
- Not large: the commit history has a single author, and there is no ecosystem of third-party integrations around it.
Side by side
| In its own words | You use it from | State between turns and runs, as documented | |
|---|---|---|---|
| agents-exe | "A handy LLM-agent tool, with a variety of calling and configuration modes" | the command line and JSON files; a Haskell library | session files; SQLite or PostgreSQL behind the server |
| LangGraph | "A low-level orchestration framework and runtime for building, managing, and deploying long-running, stateful agents." | Python; JavaScript/TypeScript | checkpointers: in-memory, SQLite, PostgreSQL |
| AutoGen | "A framework for building AI agents and applications." | Python; .NET | not checked |
| OpenAI Agents SDK | "agentic AI apps in a lightweight, easy-to-use package with very few abstractions" | Python; TypeScript | sessions, with SQLite, SQLAlchemy, Redis and other session stores in the documentation index |
| Claude Agent SDK | "Build production AI agents with Claude Code as a library" | Python; TypeScript | sessions that can be resumed or forked; storage not checked |
| CrewAI | "The leading open-source framework for orchestrating autonomous AI agents and building complex workflows." | Python | Flows "persist data across steps and executions"; storage not checked |
A few more facts from the same documentation pages, where they bear on the choice:
- LangGraph describes checkpointers as serving “conversation continuity, human-in-the-loop workflows, time travel, and fault tolerance”. That is the closest counterpart in this list to agents-exe’s stored, resumable sessions, reached from code rather than from a command line.
- AutoGen is layered: Core (“an event-driven programming framework for building scalable multi-agent AI systems”), AgentChat on top of it, and Studio, a web UI for prototyping without code.
- OpenAI Agents SDK names three primitives (agents, handoffs, guardrails) and lists built-in tracing, human-in-the-loop and MCP tools. It uses OpenAI’s Responses API by default and documents adapters for other providers.
- Claude Agent SDK is “a library that runs the Claude Code binary”, with Claude Code’s built-in tools (files, commands, web search), hooks, subagents, MCP, permissions and sessions.
- CrewAI has two layers: Crews (“teams of autonomous agents”) and Flows (“structured, event-driven workflows that manage state and control execution”). Its repository calls it “a standalone Python framework”, MIT-licensed.
When agents-exe is a reasonable pick
- You would rather describe an agent in a file and give it shell scripts as tools than write an application around a framework.
- You want the same agent in a terminal, in an MCP client and behind HTTP without rewriting it.
- You want to keep sessions in your own SQLite file or PostgreSQL database, and to run against a local model.
When to pick something else
- Your agents live inside a Python or TypeScript code base, or you need a vendor’s native features (OpenAI’s tracing and voice, Claude Code’s built-in tools and permissions).
- You want a large ecosystem of integrations, tutorials and people to ask.
- You need a hosted, managed runtime.
Sources
Read on 2026-10-03:
- LangGraph: overview, persistence, JavaScript overview.
- AutoGen: documentation home.
- OpenAI Agents SDK: Python documentation, TypeScript documentation.
- Claude Agent SDK: overview.
- CrewAI: introduction, repository.
- agents-exe: this site, the repository, and the demo for what was run.