Skip to content

Ultracode

Ultracode is Iteron's highest effort setting. It combines the maximum provider-facing reasoning intent with access to Iteron's bounded dynamic-workflow engine.

iteron --effort ultracode

Model-directed workflows

An Ultracode submission goes to the main model before any workflow is launched. The model may work directly, or call Workflow with a task-specific JavaScript program when delegation or a richer execution topology is useful. There is no mandatory stage sequence and no keyword router that turns an under-specified task into fan-out.

Scripts may compose agent(), parallel(), pipeline(), phase(), and log() in whatever bounded graph the task needs. They must handle failed agents, represented as null, explicitly. Omitting background, or setting it to false, keeps prerequisite results in the current turn; background: true is for independent work and returns a task receipt immediately.

The same engine runs standalone workflow scripts from the CLI. See iteron workflow in the CLI reference and the example script at crates/workflow/examples/repo-audit.js.

Authority remains fixed

The host, not the script, grants agent definitions and tools. Workflow calls remain bounded by the same concurrency, call, turn, token, wall-time, recursion, durable-record, and cancellation controls. Write-capable agents use host-owned worktrees; the host verifies and serially merges their patches, rejecting conflicts. A model-authored workflow cannot grant capabilities, relax a ceiling, merge its own patch, rewrite evidence, or promote a learned strategy.

Ultracode is an experimental harness policy, not an autonomous software team and not a claim of parity with another coding agent.