Tools you hand a goal — they plan, act, and decide on their own.
6 tools in this category
Scaffolding for long-horizon agents that plan, spin up sub-agents, and keep working across many steps.
Built my own deep agent without LangChain, but their tutorial is a nice starting point.
Great integration into langchain ecosystem, native langsmith support. But is rather slow latency wise, therefore weak for interactive agents. Switching to omp as a coding agent for example here resulted in less scaffolding and faster runtimes of the coding agent.
A filesystem-first framework for durable agents from Vercel. You author an agent as plain files — instructions, typed tools, skills, channels, schedules — and it persists its own state to disk, so runs survive restarts and move between environments.
Nice abstraction for durable agents, and TS tool definitions with type safety make it easy for an agent to write its own tools. Downside — like a lot of Vercel's stuff — is the tight coupling to their infra.
A terminal coding agent with the IDE wired in — language-server and debugger access (LSP/DAP), subagents, plan mode, and hindsight memory, driven by a native Rust engine. Aims to give a CLI agent the context a real IDE has.
we used this instead of deepagents at a client. It is generally more efficient, faster, better compaction or context management. Generally superseeds deepagents. There are two ways to use it! Either use it as claude code as your coding harness, or call it via omp -rpc for the autonomouses mode It works with the claude subscription, but anthropics past showed that they ban people for that as its against their tos
An open platform for software-building agents — hand it a task and it plans, writes code, runs commands, and browses the web until the work actually ships.
Openhands works really well even with small local models
Worked better than Cursor in my tests.
Andrew Ng's open-source AI coworker that lives on your desktop and returns finished work, not chat — a drafted doc, a triaged inbox, a Slack reply with the numbers. Local-first, bring-your-own-model, 25+ connectors, and it asks before anything consequential.
An agent framework from the Pydantic team that brings type-safe, validated outputs to tool-calling agents.
Evaluated it as immature a while back, but they've been cooking — now has eval, observability and a gateway in the ecosystem.