Back office ops · Production

Duolingo builds an internal fleet of coding agents using Temporal, Claude Code SDK, and Codex CLI

The problem

Duolingo engineers were spending time on repetitive coding tasks—removing deprecated feature flags, launching experiments, and modifying Terraform—rather than focusing on product thinking and core logic.

Workflow diagram · grounded in source
1
Fill out workflow JSON form
trigger
“Our Duos start by filling out a simple JSON form about the workflow. This form allows them to provide a prompt, a code repo for it to run against, and 0 or more parameters”
2
Test prompt in Codex or Claude
validation
“Duos test this prompt until they judge it successful. Generally, we take some time to craft it in Codex or Claude, and make sure it works as intended in a variety of situations”
3
Merge workflow to internal tools list
output
“Duos can easily merge their forms for them to automatically show up in a list of internal tools that any Duo can run. We also added Slack notifications to keep users informed about its progress”
4
AI agent clones repo and makes code change
ai_action
“Clone the repo The AI agent makes a code change”
5
Commit code and open PR
output
“Commit the code and optionally open a PR”
Reported outcome

Duolingo deployed a growing fleet of coding agents for routine engineering purposes; Duos can now create a simple agentic workflow in under five minutes, and average custom workflow creation takes 1-2 days.

Reported metrics
Simple workflow creation timeunder five minutes
Custom workflow average creation time1-2 days
Engineering focus shifttake routine tasks off their plates so they can focus on product thinking and core logic
Iteration speed for testing ideasFaster iteration means we can test more ideas to improve learning efficacy
Reported stack
CursorCodexClaudeCodex CLIClaude Code SDKCodingAgentTemporalSlack
Source
https://blog.duolingo.com/agentic-workflows/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Duolingo deployed a growing fleet of coding agents for routine engineering purposes; Duos can now create a simple agentic workflow in under five minutes, and average custom workflow creation takes 1-2 days.

What tools did this team use?

Cursor, Codex, Claude, Codex CLI, Claude Code SDK, CodingAgent, Temporal, Slack.

What results were reported?

Simple workflow creation time: under five minutes; Custom workflow average creation time: 1-2 days; Engineering focus shift: take routine tasks off their plates so they can focus on product thinking and core logic; Iteration speed for testing ideas: Faster iteration means we can test more ideas to improve learning efficacy (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Fill out workflow JSON form → Test prompt in Codex or Claude → Merge workflow to internal tools list → AI agent clones repo and makes code change → Commit code and open PR.

Related back office ops cases