We open-sourced AgentKit two weeks ago with zero guarantees anyone would care.
400+ clones later — we're shipping the biggest update yet. And we have benchmark data to back it up.
Quick note: AgentKit Preview is our closed, in-development intelligence layer. The fully open-source AgentKit is live and ready to use today at github.com/Ajaysable123/AgentKit — npx agentkit-ai@latest init gets you running in seconds.
Live Benchmark — Gemma 4 31b · Same Model · Same Task
Both runs used Gemma 4 31b via OpenCode. The only variable was AgentKit Preview's workflow enforcement, skill injection, and plan gates.
| Benchmark | Vanilla OpenCode | + AgentKit Preview |
|---|
| Structured planning before coding | 0% | 100% |
| Plan approved before first edit | — | ✅ Yes (40.6s review) |
| Task interruptions | 1x | 0x |
| Task completion | 20% (scaffolding only) | 80% (DER parser implemented) |
| Hard problem solved | ❌ No | ✅ Yes |
Without AgentKit — Gemma 4 31b gave up on the hard part and shipped placeholder strings ([ASN.1 Decoding Required]). No plan, no verification, interrupted once.
With AgentKit — Same Gemma 4 31b implemented a real custom ASN.1 DER parser, handled both UTCTime and GeneralizedTime, built expiration logic. Completed the task properly.
The model didn't get smarter. AgentKit's workflow gates changed its behavior:
- Plan gate forced it to think through the DER parsing approach before writing code
- Approval step made it commit to solving the hard problem instead of sidestepping it
- State machine kept it accountable through RESEARCH → PLAN → EXECUTE → REVIEW
What else just landed
Native OpenCode Integration
AgentKit now ships a native TUI plugin for OpenCode that lives inside the terminal UI — not just in the system prompt.
Select the agentkit agent from the agent switcher and you get:
- Pre-loaded skills injected automatically
- Workflow gates (RESEARCH → PLAN → EXECUTE → REVIEW → SHIP)
- Mandatory approval dialogs before any code edit
- Memory context from previous sessions
Works With Any Model
The skill router, workflow engine, and marketplace run entirely via CLI — no Claude API required. Tested on Gemma 4 31b, MiniMax M2.5, and Claude.
# Works with any model in OpenCode
agentkit workflow transition RESEARCH
agentkit workflow approve
agentkit workflow transition EXECUTE
Get started
Open-source AgentKit (free — stable & ready to use):
npx agentkit-ai@latest init
github.com/Ajaysable123/AgentKit
AgentKit Preview (closed beta — in active development)
To everyone who cloned, starred, or tried AgentKit — thank you. This is just getting started.
The callout block at the top does the heavy lifting — anyone who lands on the article immediately knows the open-source version is stable and available, and Preview is the next thing being built. Want any other changes?
