Default agents
These agent templates run the loop in default.loop.json. The orchestrator drives the iterward CLI, and the three sub-agents match the loop's steps: plan, build, and verify. Each sub-agent writes one document per dispatch, and the document's frontmatter carries its id and its status. The following table maps each file to the step that opens it, the document it writes, and the statuses the loop accepts:
| File | Agent | Step | Document | Statuses |
|---|---|---|---|---|
orchestrator.md | orchestrator | Started by the human | None | None |
planner.md | planner | plan | plan.md, and a Q-<N>.md file for each question | ok, needs-decisions, blocked |
builder.md | builder | build | build.<N>.md, which the message names | ok, blocked |
verifier.md | verifier | verify | verify.<N>.md, which the message names | pass, fail, blocked |
Each template fills in what the loop and the CLI fix: the standard message, the documents and their statuses, and the CLI commands each agent can run. Each <!-- TODO: … --> comment marks domain logic for you to write.
Each template has a Permissions section: the tools the agent uses, the files it never writes, and the shell commands it runs. Map the section to the permission settings of your agent harness, because the templates carry no harness-specific frontmatter.
Before you use the agents, replace these placeholders in every file:
<ITERWARD_DIR>: the directory of the installed package,node_modules/@push-based/iterward. Without an install, remove it from each list.<LOOP_DIR>: the directory that holdsdefault.loop.json.<RUNS_DIR>: the directory that holds one directory per run.<AGENTS_DIR>: the directory that holds these agent files.<SOURCE_DIR>: the source that the orchestrator, the planner, and the verifier never write.<BUILD_COMMAND>and<CHECK_COMMAND>: the shell commands that the builder and the verifier run.