agent harness

naïve-labs / Field guide

Agent Harness Landscape

Explore the runtimes that turn models into working agents.

32implementationsSources checked
2026-10-07

32 implementations

Coding agentAvailable

Aider

Paul Gauthier and Aider contributors

Aider works alongside you in the terminal, using repository context and Git to edit code and check changes.

Terminal · HeadlessLocalApache-2.0
Documentation
Details & sources

Useful for

Make focused changes to an existing codebase with a model of your choice.

Capabilities

  • Repository maps select relevant code context.
  • Git-aware edits and automatic commit support.
  • Linting and test feedback for iterative corrections.
Publisher’s term
AI pair programming
Models & providers
Multiple cloud providers and local models
Runs
Local
License
Apache-2.0
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Agent libraryAvailable

Browser Use

Browser Use

Browser Use supplies a Python agent loop that observes pages, chooses browser actions and checks the result.

SDK · APILocal · Self-hosted · CloudMIT
Documentation
Details & sources

Useful for

Build repeatable web tasks with a local browser or a hosted browser and agent service.

Capabilities

  • Browser state and screenshots feed iterative model decisions.
  • Custom tools and structured results in the Python library.
  • Local agent execution or a separately hosted Cloud API.
Publisher’s term
Browser agent Python library
Models & providers
Configurable model providers or Browser Use model gateway
Runs
Local · Self-hosted · Cloud
License
MIT
First available
Unknown
Announced
Unknown
Sources checked
2026-10-05

Official sources

Coding agentAvailable

Cascade

Cognition

Cascade edits code in Devin Desktop using project context, tool calls and recoverable checkpoints.

EditorLocalProprietary
Documentation
Details & sources

Useful for

Develop inside the editor while reviewing tool actions and reverting unwanted changes.

Capabilities

  • Code and chat modes with planning and terminal tools.
  • Web search, MCP connections and project rules.
  • Named checkpoints and multiple concurrent sessions.
Publisher’s term
Agentic AI assistant
Models & providers
Devin Desktop model selection; availability varies by plan
Runs
Local
License
Proprietary commercial terms
First available
Unknown
Announced
Unknown
Sources checked
2026-10-05

Official sources

Coding agentAvailable

Claude Code

Anthropic

Claude Code reads projects, edits files and runs development tools across terminal, editor, desktop and browser sessions.

Terminal · EditorLocal · Self-hosted · CloudProprietary
Documentation
Details & sources

Useful for

Build features, investigate bugs or automate repository maintenance with Claude.

Capabilities

  • Project instructions, persistent memory, skills and hooks.
  • Subagents, MCP tools and configurable permissions.
  • Programmatic access to the runtime through the Agent SDK.
  • Remote control, messaging channels and scheduled routines across supported surfaces.
  • Cloud sessions on your own machines (Team/Enterprise beta); Anthropic hosts coordination and inference.
Publisher’s term
Agentic coding tool
Models & providers
Claude through Anthropic and supported cloud-provider integrations; self-hosted cloud sessions use the Anthropic API
Runs
Local · Self-hosted · Cloud
License
Proprietary commercial terms
First available
2025-02-24
Announced
2025-02-24
Sources checked
2026-09-21

Official sources

Coding agentAvailable

Cline

Cline and contributors

Cline edits files, runs commands and uses a browser with reviewable tool actions in editors, on the desktop and in the terminal.

Editor · DesktopLocal · Self-hostedApache-2.0
Documentation
Details & sources

Useful for

Develop inside an existing editor or run the same agent engine in automated tasks.

Capabilities

  • File, command and browser tools with approval controls.
  • VS Code, JetBrains, terminal, SDK and a desktop app for macOS, Windows and Linux x64.
  • Parallel task board with separate worktrees.
  • Remote Linux workspaces over SSH and parallel subagents in the desktop runtime.
Publisher’s term
Autonomous coding agent
Models & providers
Multiple cloud providers and local model connections
Runs
Local · Self-hosted
License
Apache-2.0 for published source
First available
Unknown
Announced
Unknown
Sources checked
2026-09-21

Official sources

Relationships

Coding agentAvailable

Codex CLI

OpenAI

Codex CLI works on local projects with file and shell tools, resumable sessions, sandboxing and approval policies.

Terminal · EditorLocalApache-2.0
Documentation
Details & sources

Useful for

Implement changes interactively or run coding tasks from scripts and integrations.

Capabilities

  • Local execution with configurable sandbox and approval boundaries.
  • Project instructions, skills and MCP tools.
  • Interactive terminal and programmable execution surfaces.
Publisher’s term
Coding agent
Models & providers
OpenAI; configurable alternative and local providers
Runs
Local
License
Apache-2.0
First available
2025-04-16
Announced
2025-04-16
Sources checked
2026-10-07

Official sources

Relationships

Coding agentAvailable

Cursor Agent

Anysphere

Cursor Agent edits a project through its editor, terminal and browser tools and checks results during a task.

Editor · TerminalLocal · Self-hosted · CloudProprietary
Documentation
Details & sources

Useful for

Develop and verify changes interactively or automate the same agent through its CLI, SDK and cloud API.

Capabilities

  • Codebase search, rules and multi-file edits.
  • Shell execution and browser-based verification.
  • The same agent runtime through IDE, CLI, TypeScript or Python SDK, and cloud API.
  • Cursor-hosted VMs or tool execution on user-operated machines.
Publisher’s term
Agent
Models & providers
Cursor's supported model selection; provider and plan availability vary
Runs
Local · Self-hosted · Cloud
License
Proprietary commercial terms
First available
Unknown
Announced
Unknown
Sources checked
2026-09-21

Official sources

Agent libraryAvailable

Deep Agents

LangChain

Deep Agents supplies an extensible agent loop with file-backed context, memory and delegated subtasks for longer jobs.

SDK · TerminalLocal · Self-hosted · CloudMIT
Documentation
Details & sources

Useful for

Build a custom research or coding agent with an existing runtime and replaceable components.

Capabilities

  • Context compaction, filesystems and persistent memory.
  • Subagents, tool approvals and configurable planning.
  • Python and TypeScript libraries plus a coding CLI.
Publisher’s term
Agent harness
Models & providers
LangChain-compatible tool-calling models, including hosted and local providers
Runs
Local · Self-hosted · Cloud
License
MIT
First available
2025-07-30
Announced
2025-07-30
Sources checked
2026-09-07

Official sources

Coding agentPreview

DeepSeek Harness

DeepSeek AI

DeepSeek Harness runs coding agents through composable plugins for models, tools, sessions, sandboxes and the user interface.

Web · HeadlessLocal · Self-hostedMIT
Documentation
Details & sources

Useful for

Run or customize a local coding agent, automate one-off jobs, or embed the runtime through its SDK and ACP profiles.

Capabilities

  • Swappable Cordis plugins and configurable profiles across the runtime.
  • Web UI, headless jobs, Python SDK and ACP over standard input and output.
  • Append-only session logs for inspection, resume, fork, search and replay.
Publisher’s term
Agent harness
Models & providers
DeepSeek; Anthropic; OpenAI; Moonshot AI; Z.ai; custom OpenAI- or Anthropic-compatible endpoints
Runs
Local · Self-hosted
License
MIT
First available
2026-08-10
Announced
Unknown
Sources checked
2026-09-14

Official sources

Relationships

  • DeepSeek Harness uses components from Pi · Evidence ↗

    The generic multi-provider adapter uses Pi's pi-ai package; DeepSeek also supplies a separate direct adapter.

General agentAvailable

DeerFlow 2.0

ByteDance

DeerFlow coordinates research, coding and content tasks through subagents that share files, tools and persistent memory.

Web · TerminalLocal · Self-hostedMIT
Documentation
Details & sources

Useful for

Run longer research or creation tasks in a self-hosted workspace.

Capabilities

  • Delegated tasks with separate agent contexts.
  • Filesystem, persistent memory and sandbox backends.
  • Extensible skills, tools and an embedded Python client.
Publisher’s term
Super agent harness
Models & providers
Multiple configurable providers, including OpenAI-compatible endpoints
Runs
Local · Self-hosted
License
MIT
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Hosted agentAvailable

Devin

Cognition

Devin works through software tasks in a managed workspace with a shell, editor and browser that you can inspect or take over.

Web · MessagingCloudProprietary
Documentation
Details & sources

Useful for

Delegate bounded engineering tasks and review the resulting changes or continue the session yourself.

Capabilities

  • Managed development environment with observable tools.
  • Human takeover through the embedded editor and browser.
  • API access and messaging integrations.
Publisher’s term
AI software engineer
Models & providers
Service-managed models; arbitrary provider substitution not established
Runs
Cloud
License
Proprietary hosted service
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Coding agentAvailable

Gemini CLI

Google

Gemini CLI brings file editing, shell execution and Google Search into a terminal agent powered by Gemini.

Terminal · HeadlessLocalApache-2.0
Documentation
Details & sources

Useful for

Work on code and automate development tasks with Google's model services.

Capabilities

  • File, shell, web-fetch and search-grounding tools.
  • Project context files and conversation checkpoints.
  • MCP extensions and non-interactive scripting.
Publisher’s term
AI agent
Models & providers
Gemini through Google sign-in, Gemini API or Vertex AI
Runs
Local
License
Apache-2.0
First available
2025-06-25
Announced
2025-06-25
Sources checked
2026-09-07

Official sources

Relationships

Coding agentAvailable

GitHub Copilot CLI

GitHub / Microsoft

GitHub Copilot CLI plans, edits and tests code from a terminal, with access to GitHub workflows.

Terminal · HeadlessLocalGitHub Copilot CLI License
Documentation
Details & sources

Useful for

Work on repository tasks interactively or run bounded prompts from scripts.

Capabilities

  • Multi-step file edits and test execution.
  • Interactive terminal and non-interactive prompt mode.
  • Tool permissions, local sandbox options and GitHub task context.
Publisher’s term
Terminal-native coding agent
Models & providers
GitHub Copilot model selection; documented custom provider option
Runs
Local
License
GitHub Copilot CLI License
First available
2025-09
Announced
Unknown
Sources checked
2026-10-05

Official sources

Hosted agentAvailable

GitHub Copilot cloud agent

GitHub / Microsoft

GitHub Copilot cloud agent takes development tasks into an isolated Actions environment and prepares changes for review.

Web · EditorCloudProprietary
Documentation
Details & sources

Useful for

Delegate issues, maintenance and test improvements through GitHub's existing workflow.

Capabilities

  • Ephemeral development environments backed by GitHub Actions.
  • Code edits, tests and optional pull requests.
  • Review feedback feeds further agent work.
Publisher’s term
Cloud agent
Models & providers
GitHub-managed model selection and organization policies
Runs
Cloud
License
Proprietary hosted service
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

General agentAvailable

Goose

Goose project, Agentic AI Foundation / Linux Foundation

Goose runs local coding, research and automation tasks through a desktop app, terminal and extensible tool connections.

Desktop · TerminalLocalApache-2.0
Documentation
Details & sources

Useful for

Automate work across local tools and connected services using a chosen model provider.

Capabilities

  • Native desktop app, CLI and embeddable API.
  • MCP extensions for tools and services.
  • Cloud and local model providers.
Publisher’s term
General-purpose AI agent
Models & providers
Anthropic; OpenAI; Google; Ollama; OpenRouter; Azure; Bedrock; others
Runs
Local
License
Apache-2.0
First available
2025-01-28
Announced
2025-01-28
Sources checked
2026-09-07

Official sources

Coding agentAvailable

Grok Build

SpaceXAI

Grok Build edits code, runs commands and follows long tasks through a terminal interface with checkpoints and execution controls.

Terminal · EditorLocalApache-2.0
Documentation
Details & sources

Useful for

Interactive development, scripted maintenance and coding sessions in editors that support ACP.

Capabilities

  • Full-screen terminal, headless runs and ACP editor connection.
  • Workspace checkpoints, sandboxing and configurable tool permissions.
  • Skills, plugins, hooks and MCP tool connections.
Publisher’s term
Coding agent harness
Models & providers
Grok; configurable remote and local model endpoints
Runs
Local
License
Apache-2.0
First available
2026-05-25
Announced
2026-05-25
Sources checked
2026-09-14

Official sources

Relationships

General agentAvailable

Hermes Agent

Nous Research

Hermes Agent carries out tasks across terminal and messaging sessions while retaining memories and reusable skills.

Terminal · MessagingLocal · Self-hosted · CloudMIT
Documentation
Details & sources

Useful for

Run a persistent assistant for recurring work on your machine or a server.

Capabilities

  • Searchable past sessions and agent-maintained skills.
  • Scheduled tasks and isolated subagents.
  • Local, container, SSH and cloud execution backends.
Publisher’s term
Self-improving AI agent
Models & providers
Nous Portal; OpenRouter; OpenAI; custom and other supported endpoints
Runs
Local · Self-hosted · Cloud
License
MIT
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Coding agentAvailable

Kilo Code

Kilo (Anaconda)

Kilo Code offers coding, planning and review agents in editors, the terminal and a hosted execution service.

Terminal · EditorLocal · CloudMIT
Documentation
Details & sources

Useful for

Move between editor work and automated or cloud coding tasks with configurable models.

Capabilities

  • Specialized agents with mid-task model switching.
  • Terminal and browser tools plus MCP connections.
  • Local installations, headless CI runs and a cloud agent.
  • Resumable goals across the terminal and VS Code.
Publisher’s term
Open source coding agent
Models & providers
Multiple models through Kilo and supported provider connections
Runs
Local · Cloud
License
MIT for the current monorepo; hosted services have separate terms
First available
Unknown
Announced
Unknown
Sources checked
2026-10-07

Official sources

Relationships

Coding agentAvailable

Kimi Code CLI

Moonshot AI

Kimi Code CLI packages coding tools, parallel subagents and editor integration in a standalone terminal agent.

Terminal · EditorLocalMIT
Documentation
Details & sources

Useful for

Run Kimi coding sessions locally and migrate existing Kimi CLI conversations.

Capabilities

  • Standalone binaries, ACP integration and reusable sessions.
  • Isolated subagents, lifecycle hooks and plugins.
  • File history and automatic context compaction.
Publisher’s term
AI coding agent
Models & providers
Kimi; other compatible providers
Runs
Local
License
MIT
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

General agentAvailable

Manus

Manus AI

Manus carries out general tasks in a cloud computer with files, a browser and software tools.

Web · APICloudProprietary
Documentation
Details & sources

Useful for

Delegate multi-step research or creation tasks and inspect the resulting files and task history.

Capabilities

  • Isolated cloud sandbox with network, file system and browser.
  • Agent Mode plans and executes multi-step tasks.
  • REST API for creating tasks and retrieving results.
Publisher’s term
Agent Mode
Models & providers
Service-managed models; provider substitution not established
Runs
Cloud
License
Proprietary hosted service
First available
Unknown
Announced
Unknown
Sources checked
2026-10-05

Official sources

Coding agentAvailable

mini-SWE-agent

Princeton and Stanford research contributors

mini-SWE-agent solves coding tasks with a small shell-based loop and a linear conversation history that is easy to inspect.

Terminal · HeadlessLocal · Self-hostedMIT
Documentation
Details & sources

Useful for

Run coding experiments, evaluate models or adapt a minimal agent for local tasks.

Capabilities

  • Bash actions with independent command execution.
  • Linear histories suited to inspection and training data.
  • Local and container execution environments.
Publisher’s term
Minimal AI software engineering agent
Models & providers
Multiple model adapters, including LiteLLM and OpenRouter
Runs
Local · Self-hosted
License
MIT
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

General agentAvailable

OpenClaw

OpenClaw Foundation and contributors

OpenClaw connects models, tools and messaging channels through a persistent gateway with sessions and execution policies.

Terminal · WebLocal · Self-hostedMIT
Documentation
Details & sources

Useful for

Operate a personal or trusted-team assistant across devices and communication channels.

Capabilities

  • Persistent sessions, skills and a provider registry.
  • Tool policies and host or sandbox execution.
  • Pluggable agent runtimes behind a common gateway.
Publisher’s term
AI assistant
Models & providers
Configurable model providers and custom endpoints; optional harness plugins
Runs
Local · Self-hosted
License
MIT
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

Coding agentAvailable

OpenCode

Anomaly

OpenCode is a configurable coding agent with separate planning and editing modes and broad model-provider support.

Terminal · DesktopLocal · Self-hostedMIT
Documentation
Details & sources

Useful for

Work on a project in the terminal or connect another interface to its agent service.

Capabilities

  • Build mode, a restricted planning mode and subagents.
  • Configurable providers, tools and permissions.
  • Terminal distribution and a desktop application in beta.
Publisher’s term
Open source coding agent
Models & providers
Multiple hosted providers and compatible local models
Runs
Local · Self-hosted
License
MIT
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

Agent libraryAvailable

OpenHands Software Agent SDK

OpenHands and contributors

OpenHands provides a software-agent runtime with editable workspaces, execution tools and a server for remote conversations.

SDK · APILocal · Self-hosted · CloudMIT
Documentation
Details & sources

Useful for

Automate code changes locally or run custom agents in isolated server workspaces.

Capabilities

  • File, shell and browser tools with conversation state.
  • Local workspaces or Docker and Kubernetes execution.
  • Python, TypeScript and REST access to the same runtime.
  • Can run Pi, OpenCode and Kimi Code as optional external agents.
Publisher’s term
Software Agent SDK
Models & providers
Configurable models, including OpenAI, Claude, Qwen and Devstral
Runs
Local · Self-hosted · Cloud
License
MIT for SDK and Agent Server; cloud and enterprise terms differ
First available
Unknown
Announced
Unknown
Sources checked
2026-09-14

Official sources

Coding agentAvailable

Pi

Earendil Works; Mario Zechner and contributors

Pi combines a compact coding agent with reusable model access, agent-loop and terminal components.

Terminal · SDKLocal · Self-hostedMIT
Documentation
Details & sources

Useful for

Customize a terminal coding assistant or embed its runtime in another application.

Capabilities

  • Separate model, agent-core and coding-agent packages.
  • Multiple model providers and extensible behavior.
  • External containers or sandboxes provide isolation.
Publisher’s term
Agent harness
Models & providers
OpenAI; Anthropic; Google; other configured providers
Runs
Local · Self-hosted
License
MIT
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

  • DeepSeek Harness uses components from Pi · Evidence ↗

    The generic multi-provider adapter uses Pi's pi-ai package; DeepSeek also supplies a separate direct adapter.

  • Kimi Code CLI uses components from Pi · Evidence ↗

    Uses pi-tui terminal components; does not establish use of Pi's agent loop.

  • OpenClaw uses components from Pi · Evidence ↗

    Current dependency is pi-tui for terminal rendering; agent core is OpenClaw-owned.

Coding agentAvailable

Qwen Code

Qwen / Alibaba

Qwen Code runs coding tasks across terminal, editor and application interfaces while supporting several model protocols.

Terminal · EditorLocal · Self-hostedApache-2.0
Documentation
Details & sources

Useful for

Use Qwen or another compatible provider for interactive development and automated workflows.

Capabilities

  • Memory, skills, subagents and agent teams.
  • Multiple API protocols and local model endpoints.
  • SDKs and a shared-session daemon marked experimental.
  • Can delegate subtasks to Claude Code through Agent Client Protocol (ACP).
Publisher’s term
AI coding agent
Models & providers
Qwen, OpenAI, Anthropic and Gemini protocols; compatible local endpoints
Runs
Local · Self-hosted
License
Apache-2.0
First available
Unknown
Announced
Unknown
Sources checked
2026-09-14

Official sources

Relationships

Hosted agentAvailable

Replit Agent

Replit

Replit Agent builds and tests apps in a hosted project workspace, with checkpoints for review and recovery.

WebCloudProprietary
Documentation
Details & sources

Useful for

Create or change an app in Replit and review the generated code and running preview.

Capabilities

  • Project setup, code changes and infrastructure configuration.
  • Browser-based app testing and fix iterations.
  • Checkpoints and background tasks in the hosted workspace.
Publisher’s term
Agent
Models & providers
Replit-managed model selection; plan-dependent controls
Runs
Cloud
License
Proprietary hosted service
First available
Unknown
Announced
Unknown
Sources checked
2026-10-05

Official sources

Agent libraryAvailable

Skyvern

Skyvern

Skyvern runs browser tasks through an agent that combines page actions with multi-step goals and checks.

Web · SDKSelf-hosted · CloudAGPL-3.0 core; managed-cloud additions proprietary
Documentation
Details & sources

Useful for

Automate form filling, extraction and other browser workflows through code or a hosted interface.

Capabilities

  • Agent tasks that keep browser state across pages.
  • Playwright actions combined with natural-language actions.
  • Self-hosted server or managed browser and agent runs.
Publisher’s term
AI browser automation
Models & providers
Configurable providers in self-hosted setup; cloud options vary
Runs
Self-hosted · Cloud
License
AGPL-3.0 core; managed-cloud additions proprietary
First available
Unknown
Announced
Unknown
Sources checked
2026-10-05

Official sources

Coding agentAvailable

Zoo Code

Zoo Code Org and former Roo contributors

Zoo Code continues Roo Code's editor agent with maintained model connections, task delegation and code-search tools.

EditorLocalApache-2.0
Documentation
Details & sources

Useful for

Continue Roo-style editor workflows in the community-maintained fork.

Capabilities

  • Specialized modes and coordinated subtasks.
  • Semantic code search and MCP tools.
  • Documented Roo-to-Zoo migration path.
Publisher’s term
AI-powered dev team
Models & providers
Multiple providers including Anthropic, OpenAI, Google and Moonshot
Runs
Local
License
Apache-2.0
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

Coding agentWinding down

SWE-agent

Princeton and Stanford research contributors

SWE-agent runs configurable tool-and-feedback loops to repair software issues; its maintainers now recommend mini-SWE-agent.

Terminal · HeadlessLocal · Self-hostedMIT
Documentation
Details & sources

Useful for

Study specialized agent-computer interfaces and reproduce earlier software-agent experiments.

Capabilities

  • Configurable tools and history processing.
  • Automated repository-issue workflows.
  • Inspectable execution trajectories for research.
Publisher’s term
Software engineering agent
Models & providers
Configurable language-model providers
Runs
Local · Self-hosted
License
MIT
First available
2024
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

Coding agentRetired

Kimi CLI

Moonshot AI

Kimi CLI is an archived terminal coding agent replaced by Kimi Code CLI.

Terminal · EditorLocalApache-2.0
Documentation
Details & sources

Useful for

Inspect the former Python runtime or migrate configuration and sessions to Kimi Code CLI.

Capabilities

  • File editing, shell commands and web access.
  • Shell mode and ACP editor integration.
  • Configuration and session migration to its successor.
Publisher’s term
Terminal AI agent
Models & providers
Kimi; configured compatible providers
Runs
Local
License
Apache-2.0
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

Coding agentRetired

Roo Code

Roo Code, Inc.

Roo Code was an editor agent with specialized working modes; its extension has shut down and community development continues in Zoo Code.

EditorLocalApache-2.0
Documentation
Details & sources

Useful for

Understand earlier Roo workflows or find the documented community continuation.

Capabilities

  • Coding, planning, question and debugging modes.
  • Custom modes and MCP tools.
  • Public code remains available for inspection.
Publisher’s term
AI-powered dev team
Models & providers
Historical multi-provider support; current service availability is not assured
Runs
Local
License
Apache-2.0
First available
Unknown
Announced
Unknown
Sources checked
2026-09-07

Official sources

Relationships

Scope & method

This collection includes implemented model–action–feedback loops: coding tools, general assistants, hosted agents and libraries with a working runtime. Publisher terminology is retained. Availability describes access, not quality. Dates refer to available releases when established; announcements are separate. Documentation checks are not product benchmarks.

Coverage is strongest for coding and developer runtimes. Enterprise-only systems, browser-specific agents and regional projects need broader coverage. Proprietary internals are included only where technical details are published.

Research and curation: Silvio Haedrich / naïve-labs. Sources checked: 2026-10-07.