Delegation

Iterion can mix lightweight in-process calls and full coding-agent CLIs in one graph. backend: chooses the executor and its tool surface; model: pins the wire model independently. When backend: is absent, the normal workflow, environment, and credential resolution chain chooses one — model: alone does not force a direct API call. Delegated coding-agent backends are the richest choice for editing files, running shell, and driving git:

iter

agent implementer:
  backend: "claude_code"          # recommended (codex is supported but discouraged)
  input: plan_schema
  output: result_schema
  system: implementation_prompt
  tools: [read_file, write_file, run_command, git_diff]

Backend	Status	What it does
`claude_code`	recommended	Runs the `claude` CLI as a subprocess with full tool access
`claw`	recommended for direct provider calls	In-process multi-provider LLM client (Anthropic, OpenAI, …) with Iterion's native declared tools — select it explicitly with `backend: "claw"` and pair it with `model: "openai/gpt-5.4-mini"`, etc. It is also the last-resort backend when credential detection finds nothing.
`codex`	discouraged	Runs the `codex` CLI as a subprocess. Cannot configure its tool set, tends to fill its own context window, and has weaker iterion integration. The compiler emits a `C030` warning per node. Kept for compatibility — prefer `claude_code` or `claw`+OpenAI in new workflows.
`pi`	opt-in	Runs the pi multi-provider agent harness — the backend to reach for when you need a model the other backends cannot run (~36 providers: anthropic, openai, google, vertex, bedrock, azure, github-copilot, xai, zai, groq, cerebras, …). Reports a provider-computed cost (not an estimate), carries its own credential store, and from v3.15.0 supports the permission gate, `ask_user`/async questions, board `capabilities:`, and workflow `mcp_server:` blocks via the iterion pi extension (RPC transport only). Small tool set, no subagents/web. See Backends → Third-party agent CLIs and ADR-085.
`kimi`	opt-in	Runs Moonshot's `kimi-code` CLI, whose argv is disjoint from claude-code's (`kimi -p <prompt> --output-format stream-json [-m <alias>]`). A concrete instance of iterion's generic CLI-agent backend. See Backends → Third-party agent CLIs and ADR-065.
`grok`	opt-in	Runs xAI's `grok` Build CLI through the same CLI-agent seam (`grok -p <prompt> --output-format json [-m <model>]`); `system:` appends via `--rules`, `reasoning_effort` maps to `--reasoning-effort`, headless approval forced with `--permission-mode bypassPermissions --always-approve`. Credentials come from the CLI's own Grok Build login (`~/.grok`). Distinct from the metered xAI HTTP path (`backend: claw` + `model: "xai/…"`). See Backends → Third-party agent CLIs and ADR-065.

💡 claude_code works with your Claude subscription (Pro/Max/Team/Enterprise) — no separate API key required. claw calls provider APIs directly and needs the corresponding API key (ANTHROPIC_API_KEY, OPENAI_API_KEY, …).

Delegated coding agents are useful for autonomous implementation work. For a review, judge, or plan that should call a provider API in-process, pin both the claw backend and the model; readonly: true constrains the node independently of that selection:

iter

agent reviewer:
  backend: "claw"                      # Force an in-process provider call
  model: "anthropic/claude-sonnet-5"
  readonly: true

agent implementer:
  backend: "claude_code"               # Full coding agent — can edit files
  tools: [read_file, write_file, patch, run_command]

Leaving backend: out instead opts into automatic credential detection, which may select either claude_code or claw under the default preference.

Delegation ​

Delegation