Agent Intelligence Directory · Updated August 19, 2026
Agent Hub 2026
The platforms you run agents in, and the frameworks you build them with — what each is for, and where each stops being the right answer.
Read this before the tables
Read every capability claim here with two facts in mind. First, scaffold choice moves agentic results by up to 30 absolute points on identical models and identical tasks - the same Claude Opus 4 scored 64.9% on GAIA inside one orchestration scaffold and 57.6% inside another - so a number without its scaffold named is not a comparison. Second, an automated scanning agent has broken all eight major agent benchmarks by reward hacking, reaching near-perfect scores without solving a task. Vendor and third-party figures are recorded here as reported, labelled, and never presented as first-party measurement.
Agent platforms (6)
Products you run agents in - terminals, editors, cloud runners.
Claude Code
Anthropic
Terminal-first agent with programmable hooks and a 1M-token window over the whole repo.
OpenAI Codex
OpenAI
One account across CLI, cloud, app and mobile - async runs that open pull requests.
Cursor
Cursor
Editor and agent sharing one loop, with Composer 2.5 as the fast in-editor driver.
Google Antigravity
Free-for-individuals agent surface with multi-agent and browser task support.
Devin
Cognition
Fully autonomous runs with parallel agents, each in its own cloud VM.
GitHub Copilot (agent mode)
GitHub
The agent that already lives where the pull requests do.
Agent frameworks (6)
Libraries you build agents with - orchestration, state, handoffs.
LangGraph
LangChain
Directed graphs with conditional edges - the strongest persistence and checkpointing story.
Claude Agent SDK
Anthropic
MCP-native by design - in-process servers and lifecycle hooks rather than an adapter layer.
OpenAI Agents SDK
OpenAI
Explicit handoffs and a short path from zero to a working agent.
CrewAI
CrewAI
Role-based crews with process types, and the broadest protocol support of the group.
Pydantic AI
Pydantic
Type-first agents from the team behind the validation library everyone already uses.
Strands Agents
AWS
Simplicity-first agent building with MCP support, from the AWS side of the ecosystem.
Questions worth asking first
What is the best AI coding agent in 2026?
The question is underspecified, and that is the useful answer. Framework and harness choice alone moves agentic results by up to 30 absolute points on identical models and identical tasks, so "best agent" depends on the scaffold you run it in and the shape of your work. Pick by job: terminal-scale refactors, background pull requests, in-editor edits, or fully autonomous runs.
Why does this hub not publish a leaderboard?
Because the leaderboards are known to be gameable. An automated scanning agent broke all eight major agent benchmarks by reward hacking, reaching near-perfect scores without solving a single task. Publishing a ranked list of numbers we cannot reproduce would look authoritative and mean very little.
What is an evidence grade?
The same A-to-D scale the FrankX research hub already uses. A is reproducible first-party measurement with published receipts, B is an independent third-party benchmark, C is mixed or partly corroborated vendor figures, D is vendor-stated or editorial only. Most entries here are C or D, and saying so is the point.
How is this different from the FrankX agent catalog at /agents?
This page is the landscape: what exists in the world and what each thing is for. /agents is the proof: the 99-agent Creator OS actually built and run on this stack, packaged as installable artifacts.
Does MCP support matter when choosing?
It decides whether your tool surface travels. Frameworks built for MCP from day one inherit new protocol capabilities as they ship; frameworks that bolt it on through an adapter tend to lag. Each entry records whether MCP is native, supported, or absent.