Agent projects for workflows and tool use

Microsoft Agent Framework spans Python and .NET with graph workflows and middleware, showing how agents can sit inside a Microsoft-oriented application stack.

OpenAI Agents SDK keeps handoffs, guardrails, sessions, and tracing in a lightweight code-first toolkit, setting a simpler baseline for multi-agent coordination.

FrontierAgent combines a terminal workbench, single-agent and coordinated-team workflows, sandbox controls, session traces, and an evaluation runner in one Python project.

CrewAI separates role-based agent teams from event-driven Flows with state, routing, persistence, and human feedback, showing a Python path that combines autonomy with explicit workflow control.

Mastra places model routing, human review steps, and evals in a TypeScript framework, matching agents built into a web-oriented JavaScript application.

Koog keeps orchestration in Kotlin and Java, with Spring Boot and Ktor paths for teams that want agents inside a JVM service architecture.

AG2 v1 coordinates agents through a hub and typed channels, with optional tools and human input inside the work.

AgentScope couples agent abstractions with run visibility and fine-tuning support, widening the framework decision beyond simply wiring models to tools.

Open Multi-Agent turns a goal into a task DAG with parallel execution, plan preview and replay, approval hooks, and traces, showing coordinated handoffs beyond one agent loop.

DeepSeek Harness makes models, tools, sessions, storage, sandboxes, agent loops, orchestration, and UI replaceable through one plugin-composed agent workspace.

CLI-Anything generates and distributes structured command-line harnesses so agents can operate existing software without depending on visual clicks for every action.

TrueForge combines a chat UI, API and SDKs with model, MCP, skill, approval, session, context, and sandbox controls in a vendor-neutral runtime that can run locally or in a hosted deployment.

LiveKit Agents works through WebRTC rooms and telephony, bringing live calls and media sessions into the Python or Node.js agent comparison.

Pipecat breaks realtime audio and video into composable pipelines, transports, and flows, helping teams choose how to assemble a multimodal conversation path.

CopilotKit brings shared state, generative UI, and tool rendering into React or Angular, showing how agent actions can remain visible inside a product interface.

Dify places visual workflows, RAG, tools, and APIs between models and the finished application, testing whether non-code configuration should own that middle layer.

Hugging Face Serge places model-assisted review inside GitHub pull requests, where repository rules and human editing keep the agent output inside an existing code-review workflow.

OpenHands Agent Canvas brings backends, files, secrets, automations, and run review into one visible workspace, keeping human steering close to the work.

OpenResearch CLI keeps parallel code-and-compute experiments visible in a local dashboard, with competing directions separated in Git worktrees and run evidence tied to each experiment branch.

Orca keeps terminal-based coding agents in separate Git worktrees and brings their sessions, files, diffs, and follow-ups into one desktop review workspace.

Penpot lets compatible agents work with editable design files, components, tokens, layouts, and assets, keeping people and agents in the same collaborative design surface.

Composio packages authentication, sessions, triggers, and searchable toolkits for cases where connecting many external services could dominate the agent build.

Nango centres OAuth, API proxying, TypeScript functions, and MCP calls, putting control of auth flows and external APIs at the centre of the integration decision.

Agentic Resource Discovery defines catalogs and search for MCP servers, agents, skills, APIs, and workflows, covering the discovery step before an agent can choose or invoke an outside resource.

Google Colab CLI connects terminal and agent work to authenticated remote Colab runtimes, including compute provisioning, file execution, logs, and retrieved artifacts.

Hugging Face hf CLI gives agents a command path into Hub search, repositories, jobs, downloads, uploads, metadata, and inference services instead of requiring a browser workflow.

Mercury Agent runs as a background service with schedules, editable personality files, and permission modes, exposing the operating boundaries around recurring personal work.

Nanobot adds a packaged WebUI, goal tracking, and automation controls to a lightweight runtime, showing a durable interface beyond the terminal.

NanoClaw separates agents in containers while connecting messaging, memory, and schedules, putting task isolation ahead of a broad application framework.

Hermes Agent combines persistent memory, reusable skills, schedules, gateways, and provider switching, showing the runtime infrastructure available for a long-lived personal agent.

Atomic Agent keeps its control loop, memory, tasks, traces, and configuration on the user's machine while supporting browser, file, shell, MCP, and scheduled work.

LifeOS turns personal context, memory, skills, structured work, and a dashboard into a persistent layer around an AI coding harness, making continuity across sessions the main comparison point.

A UI-first desktop agent combines readable local memory, app connections, visual workflows, research, and agent orchestration, while keeping the managed-service boundary visible.

Headlong keeps one Bash-based agent thinking between messages, with a file-backed trajectory, shared chat stream, Docker-first setup, and explicit ongoing cost and access boundaries.

iFixAi runs fixed behavioural and governance suites with manifests and cross-provider judging, giving agent changes a repeatable diagnostic baseline.

Microsoft Agent Governance Toolkit wraps agents with policy, identity, sandboxing, and audit surfaces across several languages, making runtime controls the comparison point.

Microsoft ASSERT turns natural-language requirements into test scenarios against models or traces, making expected agent behaviour something that can be checked systematically.

Monitorability Evals supplies prompts, eval splits, mappings, and metric code for the narrower question of whether one model can monitor another agent’s behaviour.

Her Claude Code Session Detective reconstructs session traces with tool use, token use, notable actions, and turn-linked evidence, helping a reviewer diagnose what happened during a coding-agent run.

SkillSpector runs repeatable static, dependency, and optional semantic checks on agent skills, then produces terminal, JSON, Markdown, or SARIF evidence for review.

ClawWork connects professional-task artifacts, automated quality evaluation, model and search costs, and a changing benchmark balance in one agent-testing workflow.