Best AI Agent Frameworks 2026
Frameworks and platforms for building AI agents. We list 25 tools with researched pricing and features; 23 of them can be used for free.
How this list is ordered: Sorted by how much you can use for free (free or open source first, then free plan with paid tiers, then paid only), then by the number of documented features, then alphabetically. No ratings or reviews are used.
These lists are not benchmark scores; they follow the ordering above. How benchmark evidence is labeled
-
Self-organized multi-agent harness. Agents claim sub-tasks and merge progress without a central orchestrator. MIT.
- Agents discover and claim sub-tasks, share findings, verify results, and merge progress
- Runs without a central orchestrator, using a shared workspace, messages, and shared context
- Published experiments scale the harness to 1,024 agents
-
Framework that refines an agent harness on new tasks while model weights stay fixed. Apache 2.0.
- Task-level recursion proposes edits, probes them on the same task, and composes a harness
- Harnessing-level recursion keeps adopted harnesses and lessons for the next task
- Eight editable parts: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins
-
Agent harness that plans and teaches from your current understanding, then keeps a knowledge graph. AGPL-3.0.
- Measures where understanding ends, then plans a dependency map from there to the goal
- Teaches one reasoning step at a time, then runs separate practice sessions
- Stores propositions and reasoning chains in a knowledge graph across sessions
-
Safe, private runtime for fleets of autonomous AI agents. Sandboxes enforce file, syscall, and network policy. Apache 2.0.
- Runs each agent in an isolated sandbox
- Confines files and system calls with kernel controls inside the sandbox
- Checks every network connection against policy before it leaves the sandbox
-
Lets a vision model use a real iPhone via AirPlay mirroring and a $7 ESP32 acting as USB mouse and keyboard.
- Mirrors the iPhone screen over AirPlay into JPEG frames
- Controls the phone with an ESP32-S3 USB mouse and keyboard
- Uses any OpenAI-compatible vision model, cloud or local
-
Open-source framework for RAG pipelines and agents.
- Wire retrieval, reasoning, memory, and tools into agent flows
- Plug in OpenAI, Anthropic, Mistral, Hugging Face, and vector databases
- Run serializable pipelines on Kubernetes with logging and monitoring
-
Privileged access for AI agents. Vaulted credentials, per-call policy, and an audit log. Free tier is 10,000 API calls a month. Free: 10,000 calls/mo; Starter $50/mo.
- Stores credentials in a vault and injects them per request
- Checks each call against a default-deny policy
- Gives each agent its own identity and scopes
-
Open-source agent API for any model and machine. Self-host free under Apache 2.0, or use the cloud with a $0 platform fee. Platform fee $0; pay as you go for models and machines.
- Runs agents through one REST API, a TypeScript SDK, and a CLI
- Uses OpenAI, Anthropic, OpenRouter, LiteLLM, or a model you host
- Runs agents on sandboxes or on your own macOS or Linux machine
-
Python framework for agents, formerly Phidata. Paid plans from $150/month.
- Build agents, teams, and workflows using 100+ toolkits
- Expose agents as a product API or MCP server
- Run the platform in any cloud, on-prem, or locally
-
Library for letting an agent drive a browser. Credits from $5.
- Remote browsers for agents via SDK, REST, or CDP
- Hosted agents that complete browser tasks
- MIT-licensed code you can run on your machine
-
Hosted browsers for agents and automation. Paid plans from $20/month.
- Cloud browsers agents can drive to use websites
- Search for sites from a single query
- Fetch any URL as HTML, JSON, or markdown
-
Tool integrations and auth for AI agents.
- Gives agents access to more than 1,500 apps
- Loads only the catalog actions a task needs
- Manages OAuth sign-in and per-account scopes
-
Framework for multi-agent crews.
- Ranks automations from tickets, chats, apps, and workflows
- Visual editor builds workflows and exports them to Python
- Traces LLM calls, tool calls, and memory reads with costs
-
Open-source platform for LLM apps, workflows, and agents. Paid plans from $590 per workspace / year.
- Visual builder for agent workflows, prompts, tools, and decisions
- Pipelines that extract, clean, chunk, and index source files
- Create agents by chat or settings, then embed or reuse them
-
Platform for building company agents on internal data. Paid plans from 30€/seat/month.
- Connect agents to Slack, Notion, Drive, GitHub, Salesforce, and Zendesk
- Choose among 20+ models for each workflow
- Run multi-agent workflows on schedules and event triggers
-
Sandboxed cloud environments for AI agents to run code. Paid plans from $150/month.
- One Linux microVM per agent session with shell and browser
- Run sandboxes on E2B cloud or AWS, GCP, or Azure
- Run Python or JavaScript and return charts, tables, and images
-
Open-source visual builder for LLM flows. Paid plans from $35/month.
- Build multi-agent systems with workflow orchestration across agents
- Build chatbots with tool calling and retrieval from data sources
- Let humans review tasks that agents perform
-
Framework and platform for building LLM applications and agents.
- Trace agent runs step by step and search traces fast
- Turn traces into datasets and score each agent step
- Deploy agents with persistent state and isolated code sandboxes
-
Open-source tracing and evals for LLM applications. Paid plans from $29/month.
- Trace LLM calls, tool use, and retrieval steps
- Filter traces by user, session, cost, latency, or metadata
- Score outputs with model judges, rules, or human review
-
Agent runtime with persistent memory, from the MemGPT project. Paid plans from $20/month.
- Stateful agents that remember and continue learning
- Store agent memory in git-based context repositories
- Let agents reason about context during idle time
-
Framework for connecting data to language models and agents.
- Parse PDFs, Office files, spreadsheets, and images
- Extract tables, charts, and handwritten text into structure
- Pull defined fields from documents without model training
-
TypeScript framework for agents, workflows, and RAG. Paid plans from $250/month.
- Define typed agents with instructions, models, and tools
- Compose multi-step workflows with retries and branches
- Store context using observational memory and semantic recall
-
Serverless cloud for running model and agent code. Paid plans from $250/month.
- Serve language, image, video, and audio models live or in batch
- Run isolated sandboxes for coding agents and untrusted code
- Fine-tune models and run reinforcement learning on GPUs
-
Runs your agents in isolated sandboxes on Docker-managed cloud infrastructure. Experimental. First $25 of compute included; then per-second rates.
- Predefined sandboxes for Claude Code, Codex, Gemini CLI, Copilot, and OpenCode
- Stores model API keys as secrets, never inside the sandbox
- Network policies decide what each sandbox can reach
-
Describe a change in plain language. The agent scopes it across your indexed codebase and tracks it through to merge. Pay per changeset merged.
- Scopes the change across your indexed codebase
- Applies the change and tracks each changeset through to merge
- On by default for Sourcegraph Cloud customers with AI features
