No-Code Agentic Workflow Platform for Automated Code Changes

duolingo · Education · 2025 · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Internal AI coding automation

Duolingo needed to scale AI-assisted code changes beyond individual engineers by letting employees create and deploy coding agents without custom code.

Constraints:

Employees across roles needed to define agents without writing custom orchestration code.
The platform had to support multiple LLM providers.
Routine tasks include feature flag removal, experiment management, and infrastructure changes.
Workflows needed durable orchestration for multi-step code-change tasks.

Architecture

Duolingo built a JSON-based workflow creator backed by a unified CodingAgent library for Codex and Claude and orchestrated executions through Temporal workflows.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  json-workflow-creator["JSON Workflow Creator"]
  codingagent-library["CodingAgent Library"]
  llm-providers["LLM Providers"]
  temporal-workflows["Temporal Workflows"]
  target-repositories["Target Repositories"]
  json-workflow-creator --> codingagent-library
  codingagent-library --> temporal-workflows
  codingagent-library --> llm-providers
  llm-providers --> codingagent-library
  temporal-workflows --> target-repositories

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`JSON Workflow Creator`	Lets users define prompts, repositories, and parameters without custom code.	`CodingAgent Library`
`CodingAgent Library`	Provides a unified interface over Codex and Claude coding agents.	`Temporal Workflows`, `LLM Providers`
`LLM Providers`	Execute coding-agent tasks through Codex and Claude.	`CodingAgent Library`
`Temporal Workflows`	Durably orchestrate multi-step agent runs.	`Target Repositories`
`Target Repositories`	Receive automated code changes for routine tasks.	—

Data flow

A user defines prompts, repositories, and parameters in the workflow creator; the CodingAgent library routes to Codex or Claude; Temporal orchestrates durable multi-step execution; and the agent produces code changes for routine engineering tasks.

“The solution centers on a JSON-based workflow creator that allows users to define prompts, target repositories, and parameters, backed by a unified CodingAgent library supporting multiple LLM providers (Codex and Claude) and orchestrated through Temporal workflows.”summary, stated

“The platform has enabled rapid creation of agents for routine tasks like feature flag removal, experiment management, and infrastructure changes”summary, stated

Why this architecture

The platform separates workflow definition from provider-specific agent execution so non-specialists can create repeatable coding automations while Temporal handles durable orchestration.

“enabling employees across all roles to create and deploy AI coding agents without writing custom code”summary, stated

“custom multi-step workflows buildable in 1-2 days”summary, stated

Models used: Codex · Claude · Temporal

Scale

Simple agents can be deployed in under five minutes.
“simple agents deployable in under five minutes”summary, stated
Custom multi-step workflows can be built in 1-2 days.
“custom multi-step workflows buildable in 1-2 days”summary, stated

Alternatives considered (roads not taken)

Individual engineers writing bespoke coding-agent scripts

Rejected because: Duolingo wanted employees across all roles to create agents without custom code.

“without writing custom code”summary, stated

Single-provider coding-agent platform

Rejected because: The library needed to support multiple LLM providers including Codex and Claude.

“supporting multiple LLM providers (Codex and Claude)”summary, stated

Patterns

`tool-use-orchestration` (agent-architecture)

Coding agents are configured to operate over target repositories and routine engineering workflows.

“define prompts, target repositories, and parameters”summary, stated

`durable-execution` (workflow)

Temporal workflows orchestrate multi-step coding-agent runs.

“orchestrated through Temporal workflows.”summary, stated

`model-routing` (model-selection)

A unified CodingAgent library supports Codex and Claude behind one workflow interface.

“supporting multiple LLM providers (Codex and Claude)”summary, stated

Tradeoffs

Use JSON workflow definitions

Chose No-code workflow creator over Custom code for each agent.

Rationale: This made routine coding agents accessible across roles.

“JSON-based workflow creator that allows users to define prompts, target repositories, and parameters”summary, stated

Abstract over multiple providers

Chose Unified CodingAgent library over Provider-specific workflow code.

Rationale: The platform needed to support both Codex and Claude while presenting one workflow model.

“backed by a unified CodingAgent library supporting multiple LLM providers (Codex and Claude)”summary, stated

Gotchas (what bit them)

Scaling coding agents requires a productized workflow layer, not only individual agent use.

Root cause: Individual AI-assisted development does not let all roles create repeatable automations.

Resolution: Expose a JSON workflow creator with prompt, repository, and parameter controls.

“scaling AI-assisted development beyond individual use.”summary, stated

Multi-step code-change workflows need durable orchestration.

Root cause: Routine code changes can span repositories, prompts, and execution steps.

Resolution: Run agent workflows through Temporal.

“orchestrated through Temporal workflows.”summary, stated

Evaluation

Duolingo reports agent creation/deployment times and workflow-build times as operational productivity metrics.

Metrics: Simple agent deployment time · Custom workflow build time · Routine coding task coverage

“simple agents deployable in under five minutes and custom multi-step workflows buildable in 1-2 days”summary, stated

Outcomes

Simple agents can be deployed in under five minutes.
Custom multi-step workflows can be built in 1-2 days.
Agents cover feature flag removal, experiment management, and infrastructure changes.

The platform is internal and cross-role; the case reports creation speed rather than total agent count.

“allowing engineers to focus on core product logic rather than repetitive coding tasks.”summary, stated