AI Agents List [2026] | Frameworks, Agentic Harnesses & Useful Repos

Community Article
Published September 24, 2026

Apparently quite a few of you found my 2025 agents list useful, so here we go again :D

image/jpeg

The original was basically my own notepad that I decided to post before losing another version somewhere on my desktop lol. This one keeps that idea, with a much bigger list and a bit more organisation.

There are frameworks for building your own agents, coding tools you can actually open and use, research agents, browser automation, voice agents, and a dedicated section for agentic harnesses. Also included the useful boring stuff: testing, tracing, permissions and checking whether a project is still maintained.

Last Updated: 24 September 2026.

This is a research directory, not a benchmark or a claim that every project has been personally tested. The main entries have public documentation and recent repository or release activity. Smaller projects and slower moving tools are labelled separately. I do my best to actually read through community feedback and real world user write ups when I research and create these lists / knowledge bases etc but as always, do your own research and go in with an open mind when you test out stuff :D

....that's what keeps it fun (for me anyway haha!)

Love always, it's a crazy world atm and stuff is moving so quickly, be kind to eachother and share knowledge and we must just get through this all good <3

~tegridydev

Most entries are open source. Tools with commercial dependencies or extra licence restrictions have their own section, with exceptions called out beside the relevant project.

Find what you need

Quick picks

My starting shortlist based on the documentation, maintenance signals and use cases reviewed for this update:

What you want to do Start looking here
Build a Python agent with typed inputs and outputs Pydantic AI
Control a workflow with branches, checkpoints and human review LangGraph
Learn how agents execute generated code smolagents
Build an agent into a TypeScript application Mastra or VoltAgent
Build within a Microsoft Python or .NET stack Microsoft Agent Framework
Use a coding agent in your terminal Codex CLI or OpenCode
Customise a small, extensible harness Pi, with isolation configured separately
Build your own longer running agent Deep Agents, Pydantic AI Harness or Strands
Try an assistant with memory and scheduled tasks Hermes Agent
Produce research reports with source links GPT Researcher
Automate browser tasks Browser Use or Stagehand
Build a voice assistant LiveKit Agents or Pipecat

Pick the section that matches your actual job. You do not need to install the whole list.

Frameworks and SDKs

These are the building blocks for writing your own application. Expect to define tools, configure models and decide what happens when something fails.

  • Pydantic AI: Python agents with typed tools, validated outputs and multiple model providers. A sensible starting point if your agent needs to return data your application can reliably consume. MIT.

  • LangGraph: Stateful workflows with branching, persistence, checkpoints and human intervention. Worth looking at when a task needs to pause, resume or recover. You get more explicit control, with more workflow design to do. MIT.

  • smolagents: Hugging Face's compact agent library, particularly useful for agents that write code to perform actions. Good for learning and experiments. Configure an appropriate execution sandbox before running generated code. Apache 2.0.

  • CrewAI: Organise agents around roles and tasks, with Flows for controlling the wider process. Useful for exploring specialist teams. Start with a small crew so you can see whether each extra agent earns its token bill. MIT core.

  • Agno: Build agents, teams and workflows, then expose and manage them through its AgentOS runtime. Includes storage, approvals and observability. Review the documented telemetry settings when configuring a private deployment. Apache 2.0.

  • OpenAI Agents SDK: Agents, tools, handoffs, sessions, guardrails and tracing. A relatively direct way to build agent workflows; the project also supports other model providers. There is a JavaScript and TypeScript SDK too. MIT.

  • Google ADK: Google's Agent Development Kit for building, evaluating and deploying agents. Useful if you want an SDK with workflow patterns and development tooling. Check which features depend on your chosen model or deployment service. Apache 2.0.

  • Microsoft Agent Framework: Agents and orchestration for Python and .NET, including workflow state, approvals and telemetry. This is the starting point Microsoft's AutoGen repository now recommends for new projects. MIT.

  • Mastra: TypeScript agents with workflows, memory, tools, evaluation and application integrations. A useful option for a web application stack. Apache 2.0 core; directories named ee/ have separate enterprise terms.

  • VoltAgent: TypeScript framework with tools, memory, workflows and agent supervision. Worth comparing with Mastra if you want to keep application logic in TypeScript. The VoltOps platform is a separate product surface to check. MIT framework.

  • Haystack: Python pipelines and agents with explicit retrieval, routing and generation components. Especially relevant when your agent needs to work with documents and search. Apache 2.0.

  • LlamaIndex: An established toolkit for connecting documents and data to RAG and agent applications. The current README says the company's focus has shifted towards parsing and extraction. Check that roadmap against what you need from the broader framework. MIT framework; hosted products have separate terms.

  • AG2: Agent collaboration, tool use and human participation, with its own active development path. Treat AG2 and Microsoft's AutoGen as distinct projects when following documentation or installing packages. Apache 2.0.

  • CAMEL: Agent collaboration, simulations, tool use and synthetic data generation. Particularly interesting if you want to research how agents behave together. Apache 2.0.

  • DSPy: Build and optimise modular language model programs, including agent loops. Useful when you have examples and an evaluation metric to optimise against. You still need to define what a good result actually looks like. MIT.

  • Rig: Rust library for model integrations, tools, retrieval and agentic workflows. An option for keeping your application in Rust. The maintainers flag breaking changes, so pin versions and read migration notes. MIT.

Agentic harnesses and coding agents

The model generates the next response or action. The harness is the software around it: the execution loop, tools, context handling, sessions, errors and whatever permission controls the project implements.

Think of the model as the engine. The harness supplies the controls and connects it to the rest of the machine. Some harnesses leave quite a lot of those controls for you to configure.

This category overlaps with frameworks. Some entries are ready to use coding tools; others are libraries for assembling your own agent.

Tools you can open and use

  • Codex CLI: OpenAI's local coding agent for working inside a repository. Useful for changes that involve reading files, editing code and running checks. Review its sandbox and approval configuration. Apache 2.0 CLI; model and service terms are separate.

  • OpenCode: Coding agent with a terminal interface, desktop option and provider flexibility. A good candidate if you want to try different models in one tool. Its permission documentation says most permissions default to allowing actions, so configure these deliberately. MIT.

  • Pi: Extensible terminal harness with an agent runtime and model API packages. Interesting if you want to shape the tool around your workflow. It does not include a built in system for restricting filesystem, process, network or credential access. Use the documented isolation options. MIT.

  • goose: General assistant with desktop, terminal and API interfaces, multiple providers and MCP extensions. Covers coding, research and automation. The current repository is under aaif-goose; older block/goose links redirect there. Apache 2.0.

  • Gemini CLI: Google's terminal agent with file operations, shell tools, search and MCP support. An obvious candidate if you want to work with Gemini. Check the current account terms and limits instead of relying on an old free tier screenshot. Apache 2.0.

  • Qwen Code: Coding agent with multiple provider protocols, local model support and interfaces beyond the terminal. Worth checking for a Qwen or local model setup. Stable and nightly releases are separate choices. Apache 2.0.

  • Cline: Coding agent available through an editor, CLI, desktop app and SDK. Its editor workflow includes action approvals. The repository is Apache 2.0, but its README says the JetBrains plugins are not currently open sourced. See the security note below about the February CLI package incident.

  • Kilo Code: Coding agent across editors and the terminal, with model switching and a broader platform around it. Useful if you want several ways to access the same workflow. Check the distinction between repository code and hosted services. MIT repository.

  • OpenHands: The current main repository centres on Agent Canvas, a control interface for coding agents and automations across local, remote and cloud backends. Check current instructions because older OpenHands tutorials describe a different setup. MIT repository; hosted services are separate.

  • mini SWE agent: A deliberately small software engineering agent built around shell actions. Useful for understanding the loop, experiments and reproducible evaluations. Choose an isolated execution environment before giving it a repository. MIT.

Harnesses to build on

  • Deep Agents: A harness built on LangGraph with planning, filesystem tools, subagents, context management and persistence. Useful when you want those pieces assembled before adding your own tools and workflow. MIT.

  • Pydantic AI Harness: Adds capabilities such as planning, memory, workspace tools and subagents to Pydantic AI. A natural next step if you already use that framework. The current Coder README explicitly describes unrestricted shell commands; that needs a deployment boundary. MIT.

  • Strands Agents: SDKs and assembled harnesses for Python and TypeScript, with lifecycle controls, tools, sessions, memory and evaluation support. Useful for owning the agent loop inside your application. Apache 2.0.

Before choosing a harness, compare the things you will notice after the demo: can you inspect a failed run, stop it, control its tools, resume work and measure what it cost?

Personal assistants, research and browser agents

  • Hermes Agent: Nous Research's assistant with memory, reusable skills, scheduled tasks and messaging integrations. Interesting for ongoing work across sessions. Review the execution backend and what its saved memories and skills contain. MIT.

  • OpenClaw: Assistant that connects to messaging platforms and tools while keeping its gateway and state on your infrastructure. Its security model assumes a trusted operator boundary. Main session tools run on the host unless sandboxing is configured. MIT.

  • Letta Code: Stateful agent harness with persistent memory, skills and interactive or ongoing operation. The older Letta repository now points here for current development; the V1 server is historical. Useful for exploring continuity across tasks. Apache 2.0.

  • Agent Zero: Agent environment with a Dockerised Linux desktop, browser, projects, memory and delegation. Useful for tasks involving a whole workspace. Its host bridge and mounted folders determine how much of your actual machine it can reach. MIT.

  • GPT Researcher: Research agent for gathering web or local information and producing reports with citations. A useful foundation for research workflows. Open the citations and check they support the claims before sharing the report. Apache 2.0.

  • DeerFlow: ByteDance's harness for extended research, coding and creation tasks, using skills, subagents, memory and sandboxes. Version 2 is a rewrite; tutorials for the original research framework may not apply. MIT.

  • Browser Use: Browser agents for navigating sites and completing tasks. The open library and hosted browser services are different deployment choices. Start with a separate browser profile and a test account. MIT library.

  • Stagehand: Browser SDK for actions and structured extraction, with direct browser control alongside model assistance. Useful when you want to combine explicit automation with natural language steps. MIT SDK; Browserbase hosting is separate.

Voice agents

  • LiveKit Agents: Framework for realtime voice and multimodal agents, with speech services, turn detection, WebRTC and telephony integration. Useful for assistants people speak to through an app or phone. Apache 2.0.

  • Pipecat: Python framework for streaming voice and multimodal pipelines. Useful when you want to compose speech recognition, models, speech generation and transports yourself. BSD 2 Clause.

Visual builders

  • Langflow: Visual authoring for agents and workflows with Python components, API access and MCP integration. Handy for exploring a flow before embedding it into a larger application. MIT.

  • Sim: Workspace for building, deploying and monitoring agents and workflows, with integrations and self hosting support. An option for teams that want a shared interface around their automation. Apache 2.0 repository.

Sandboxes, tracing and testing

These support agents. They are not standalone assistants.

  • E2B: SDKs and infrastructure for running generated code in isolated environments. Useful when your agent needs to execute Python, shell commands or other tools. Check hosting, network access and credential handling for your deployment. Apache 2.0 repository.

  • Langfuse: Tracing, prompt management, datasets and evaluations. Useful for seeing what your agent actually did and comparing changes. MIT outside the separately licensed ee directories. Keep sensitive content out of traces where possible.

  • Promptfoo: Evaluation and adversarial testing for prompts, agents and RAG applications. Useful for regression checks and testing failure cases before deployment. Passing a test suite is evidence about those tests, not proof that an agent is secure. MIT.

Protocols worth knowing

Project What it connects Useful distinction
Model Context Protocol AI applications with tools and data The official server repository contains reference implementations. Its README explicitly says they are educational examples, not production ready solutions.
Agent Client Protocol Editors and clients with coding agents Helps separate your preferred editor from the agent you run.
Agent2Agent Separate agent applications Supports discovery and communication between systems without requiring them to expose their internals.

Protocol support does not tell you whether a particular plugin is trustworthy. Inspect the actual server, package and permissions you are about to use.

Commercial dependencies and licence exceptions

These belong in a practical directory, with the licensing made clear.

Project What it does What to know
Claude Code Coding agent for terminal and editor workflows Proprietary product. The public repository is not an open source release of the full runtime. Its licence points to Anthropic's commercial terms.
Claude Agent SDK Build applications using the Claude Code agent runtime The Python SDK is MIT licensed, but bundles and depends on the separately governed Claude Code CLI.
Dify Visual agent workflows, RAG and application development Uses a modified Apache licence with additional conditions, including restrictions concerning multiple tenants and frontend branding. Read its actual licence.
n8n Business automation with integrations, workflows and agent steps Source available under its Sustainable Use and enterprise licences. Do not describe it as ordinary permissive open source.
AutoGPT Build and run agents through a workflow platform autogpt_platform/ uses Polyform Shield; Classic and the remaining code have MIT licensing. The old standalone agent and current platform are different things.
GitLab Duo Agent Platform Agents and flows integrated with GitLab's development workflow A commercial platform with version, offering and entitlement requirements. Public GitLab source does not make all Duo functionality freely licensed or free to run.

These are short labels to help you compare projects. The licence shipped with the version and component you use is what matters.

More agent lists and examples

If this still has not given you enough tabs:

  • Awesome LLM Apps: Examples covering agents, RAG applications, skills and practical workflows. Good for finding an implementation to read. Review each example's dependencies before running it.
  • Awesome AI Agents: E2B's broad directory of agents and related products. Contains both open and closed offerings, including older projects.
  • Awesome Agents: Frameworks, research, automation, memory and other agent tooling. Useful for discovery; check the linked project's current status.
  • Awesome ADE: Agentic development environments, orchestration interfaces, terminal tools and isolation options. Particularly relevant if you are building a workspace around coding agents.
  • Awesome AI Agent Papers: A 2026 paper collection covering agent engineering, memory, evaluation and workflows. Treat paper claims and working software as separate things to assess.

An entry in an awesome list is a starting point for research, not a security review.

misc tools

  • Ferrum (official GitHub mirror): Small Rust coding agent for Linux, with file and shell tools, sessions and configurable providers. The mirror describes it as an active beta. MIT. Emerging project; independent user feedback is limited.
  • Botfile (GitHub mirror): Manages shared instructions and skills across agent tools using symlinks. Useful supporting tooling, rather than an agent itself. Apache 2.0. Mirror activity is older; current upstream maintenance remains unverified here.

Two established names also need a maintenance caveat:

  • Aider: Terminal pair programming with repository maps and Git integration. Still relevant to the history and design of coding tools, but the reviewed release page remains at v0.86.0 and the repository's latest reported push was 22 May 2026. Apache 2.0. Slower activity; not presented as a freshly maintained default.
  • AgentOps: Agent monitoring, replay and cost tracking. MIT. The repository's latest reported push was 25 June 2026. Check maintenance and compatibility with your current SDK versions before adopting it.

Projects whose status has changed

Some familiar names now need a different label.

Project Current note
Microsoft AutoGen Maintenance mode. Its README directs new users to Microsoft Agent Framework.
Continue The README says the repository is no longer actively maintained and describes a final 2.0.0 release.
Flowise Archived. Its README points to a discussion about the project's future.
Roo Code Archived. Its README says the extension shut down on 15 May and points users towards alternatives.

What community feedback actually helped with

The useful discussions i found were about everyday friction: context overhead, model compatibility, debugging, token costs and how much configuration people could tolerate.

In this LocalLLaMA harness discussion, several users liked Pi's simplicity and extensibility. Others preferred OpenCode, Codex, Cline or Hermes. There were also complaints about extra tools and plugins making local workflows more complicated.

Local agent comparisons and production framework discussions are useful places to look for questions to test yourself. They are anecdotal, include different setups and may include people promoting their own work.

My takeaway: test your own task with your own model. Measure completion, mistakes, cost and recovery. A long list of features will not tell you those things.

A few things before giving an agent your terminal

Open source means you can inspect the code. It does not mean the software has been audited or that its defaults suit your machine.

  • Start in a disposable workspace with a clean Git checkpoint.
  • Configure filesystem, shell and network boundaries. An approval prompt and an operating system sandbox do different jobs.
  • Give tools only the credentials and account access they need. Keep production secrets away from experiments.
  • Treat third party skills, plugins and MCP servers as software you are installing, even when the install takes one command.
  • Set limits on runtime, retries and spend. Keep approvals for consequential actions such as sending, deleting, publishing or buying.
  • Check releases and security advisories before installing. Pin a reviewed version and keep a way to roll back.

A concrete example: Cline disclosed an unauthorised CLI package publication on 17 February 2026. The advisory identifies the affected 2.3.0 package, a corrected release and changes to publishing. That is a reason to check the exact package and version, rather than assume a popular repository makes every release trustworthy.

Pi's permission and containerisation notes, OpenCode's permission defaults, and OpenClaw's security guide are worth reading before connecting those tools to anything important.

If you have a useful project to add, drop the repo, what it actually does, the licence and a recent release or maintenance link in the comments. Bonus points for something you have used and can explain the rough edges of.

Follow, bookmark, save the tab. Hopefully this version survives somewhere more sensible than my desktop :D

ty for reading <3 | tegridydev.com

Community

Sign up or log in to comment