Workflow Authoring

Workflow authoring reference

A workflow structures work across many subagents — to be comprehensive (decompose and cover in parallel), to be confident (independent perspectives and adversarial checks before committing), or to take on scale one context can't hold (migrations, audits, broad sweeps). The script is where you encode that structure: what fans out, what verifies, what synthesizes.

When you do call the Workflow tool, the right move is often hybrid: scout inline first (list the files, scope the diff, find the modules) to discover the work-list, then call Workflow to pipeline over it, passing the list via args. You don't need to know the shape before the task — only before the orchestration step.

Common single-phase workflows you can chain across turns:

• Understand — parallel readers over relevant subsystems → structured map
• Design — judge panel of N independent approaches → scored synthesis
• Review — dimensions → find → adversarially verify
• Research — multi-modal sweep → deep-read → synthesize
• Migrate — discover sites → transform each → verify

For larger work, run several in sequence — read each result before deciding the next phase. You stay in the loop; each workflow is one well-scoped fan-out.

Every script must begin with export const meta = {...}:

export const meta = {
  name: 'find-flaky-tests',                       // kebab-case, required
  description: 'Find flaky tests and propose fixes',  // required
  phases: [                                       // one entry per phase() call
    { title: 'Scan', detail: 'grep test logs for retries' },
    { title: 'Fix', detail: 'one agent per flaky test' },
  ],
}
// script body starts here — use agent()/parallel()/pipeline()/phase()/log()

The meta object must be a PURE LITERAL — no variables, function calls, spreads, or template interpolation.

What a subagent is

Every agent() call is one agent-free ephemeral conversation linked to the invoking parent agent (is_subagent: true, named after the call's label). It starts with nothing but its prompt: no memory, no conversation context, no skills. Put ALL context a stage needs in the prompt — file paths, the rule it should apply, what shape to return.

Subagents are told their final text IS the return value (not a human-facing message), so they return raw data. Workers inherit the invoking session's tools and permission mode; use allowedTools to restrict a workflow or stage. For stages whose input is entirely in the prompt (synthesis, judging, scoring) pass allowedTools: [] — a model that can still read files tends to wander, and a subagent that re-issues an identical tool call three times is stopped and rejects with the failure detail.

Their model defaults to the invoking conversation's model. opts.model (or the tool's model input) accepts any handle or alias listed by letta model list; an unknown value rejects that call. Use a cheaper model for mechanical stages only when you know a valid handle.

Use the invoking backend for workflow workers. Local execution requires an Agent SDK and App Server with local conversation support; decide() still uses the Cloud decisions service.

Script body hooks

• agent(prompt, opts?) → Promise. Spawn one subagent. Resolves to its final text, or with schema (JSON Schema) to a validated object. Prefer schema for shaped results: invalid or missing output is retried, then rejects with validation detail in the error and journal. json: true still parses without validating; schema wins if both are set. Await each call; failures reject with the cause, callIndex, and conversationId when available. Catch errors explicitly to recover. Options: label (display name), phase (progress group — use this inside concurrent stages), schema, json, model, effort ('low' for mechanical stages, higher for the hardest verify/judge stages), allowedTools, systemPrompt (extra system prompt for this subagent), timeoutMs (default 10 minutes), maxToolCalls (positive safe integer; default 1000 unique tool calls for this subagent), conversationId (resume a worker — see "Diagnosing a run").

• pipeline(items, stage1, stage2, ...) → run each item through all stages independently, NO barrier between stages. Item A can be in stage 3 while item B is still in stage 1. This is the DEFAULT for multi-stage work. Wall-clock = slowest single-item chain, not sum-of-slowest-per-stage. Every stage callback receives (prevResult, originalItem, index) — use originalItem/index in later stages to label work without threading context through stage 1's return value. A stage that throws skips that item's remaining stages; siblings finish, then the helper rejects.

• parallel(thunks) → run zero-arg functions concurrently. This is a BARRIER: it awaits all thunks, then rejects if any threw. For best-effort processing, catch errors inside each thunk and return explicit success/error results. Use ONLY when you genuinely need all results together.

• phase(title) — start a new phase; subsequent agent() calls are grouped under this title in progress output.

• log(message) — emit a progress message to the user.

• decide(state, questions, opts?) → Promise; not a subagent call. Ask a calibrated Jev model (chosen internally — no model option) questions about state. questions is a non-empty OBJECT keyed by question id — never an array. Each question needs instructions and a type: choice (criteria map of id → description, ≤255), score (criteria array of strings, unlike questions), noul (no criteria). Returns a response whose answers are keyed by the same ids and calibrated, or null after one retried invalid answer — guard if (!call); API errors throw.

  const call = await decide(evidence, {
    behavior: { type: 'choice', instructions: 'Is this behavior a bug?',
      criteria: { bad: 'Wrong or harmful', not_bad: 'Expected or harmless' } },
  })
  const verdict = call?.answers?.behavior?.choice  // 'bad' | 'not_bad'
  

• tools.mcp__server__tool(args?) → Promise; not a subagent call. Calls one of the invoking agent's MCP tools directly (find names with letta mcp search / letta mcp tools before writing the script). Each call goes through the same permission rules as a normal MCP tool call; a tool that would ask for approval throws instead, so the user must allow it first. Resolves to the tool's structuredContent, else its text output (parsed when it is JSON), else the raw content array. Throws when the tool is unavailable, not allowed, or reports an error; parallel() and pipeline() propagate that rejection. Use it for deterministic fetches and writes whose arguments the script already knows; use agent() when a step needs judgment.

  const issues = await tools.mcp__linear__list_issues({ team: 'LET', limit: 20 })
  

• args — the value passed as the tool's args input, verbatim. Pass arrays/objects as actual JSON values, NOT as a JSON-encoded string.

decide() sees only the state you pass; it cannot read files. When compact, bounded items are already prepared, pass them via args and call decide() on each directly — don't spawn agent() readers just to relay inputs. Raw large traces don't belong in args: prepare bounded state that preserves the user instruction, observed action, and outcome, and mark what was omitted.

Scripts are plain JavaScript, NOT TypeScript — type annotations, interfaces, and generics fail to parse. The script body runs in an async context — use await directly and return the final result. Standard JS built-ins (JSON, Math, Array, etc.) are available; the hooks are the only globals provided. The script runs inside the CLI process with the CLI's own privileges (the vm context is a scope, not a security boundary), and the user approves it by reading it. Keep the script to orchestration: decide what runs and combine results. Direct MCP calls belong in tools.mcp__*(); other reading, searching, and writing belongs in subagents, where the tool allowlist applies.