FlowEngine: Orchestrate Multi-Step AI Workflows

Today we're launching FlowEngine — a production-grade workflow orchestration system built directly into Poly. If you've ever found yourself copying outputs from one AI conversation into another, chaining prompts manually, or wishing you could automate a research process you run every week, FlowEngine is for you.

How it works

FlowEngine models every workflow as a directed acyclic graph. Each node is a step — search the web, analyze sources, verify claims, synthesize findings, generate a report. Edges define dependencies and data flow. The engine handles scheduling, parallelism, retries, and timeout enforcement automatically.

You define workflows in YAML. Here's a minimal example:

name: quick-research
steps:
  - id: search
    kind: search_web
    params:
      query: "{{input.topic}}"
  - id: synthesize
    kind: synthesize
    depends_on: [search]
    params:
      sources: "{{search.results}}"

Thermal orchestration

Thermal orchestration keeps things running smoothly under load. When the system is cool, steps execute at full speed. Under heavy load, non-critical steps are deprioritized. At critical temperatures, the engine sheds optional work to protect your core workflow. You get results — always.

The four thermal tiers are: cool (full speed), warm (slight backoff), hot (non-critical steps deprioritized), and critical (emergency throttle). The orchestrator measures system load continuously and transitions between tiers automatically.

59 templates

We're shipping with 59 templates covering research, code analysis, financial reporting, content generation, and more. Each template is a YAML file you can customize or use as a starting point for your own workflows. The template validator catches structural issues before execution — no more debugging DAGs in production.