Executors

An executor is what actually runs an agent node. Azhi does not hide the differences between them. Each one declares its capabilities, and the run plan checks the workflow against them.

The executors

Executor	Providers	Notes
`model-agent` (default)	Anthropic, OpenAI	A platform-owned tool loop. Every tool call goes through the gateway; no ambient tools exist
`opencode`	Anthropic, GitHub Copilot, ChatGPT plan	Gateway bridged over MCP, built-in tools restricted by deny-all permissions, isolated `HOME`
`claude-agent-sdk`	Anthropic	The SDK driving the Claude Code runtime. Verified against a scripted endpoint; not yet against a live model
`codex`	OpenAI	Experimental, off by default

What each declares

Capability	Values
streaming	yes / no
resume	native / checkpoint-only / none
cancellation	confirmed / best-effort / none
gateway tools	native-mcp / bridged / none
ambient tools	disableable / restrictable / uncontrolled
structured output	enforced / validated / none
usage	reported / partial / unavailable
platforms	linux, macos, windows
compacts context	yes / no. If yes, the platform does not compact

Anything not yet confirmed by a conformance run is listed as unverified and the plan says so. For example, OpenCode’s usage, structured output, cancellation and platform support are declared unverified.

Why bring your own harness?

OpenCode and the Claude Agent SDK are strong interactive tools. Azhi runs them as steps and adds what they leave to you: durability, scheduling, approvals, the action ledger, taint checks and cost limits.

Isolated coding workspaces

An agent step can get a fresh git checkout on a worker, with isolated git and HOME, deleted after the step. The checkout token is read-only and never written. At most 2 runs work on one repository at once by default.

Compare them

npm run compare runs 30 fixtures through two executors and writes docs/executor-comparison.md. It states which numbers are estimated and that OpenCode’s usage is incomplete.