Stacks/Coding agent

Duolingo AI STR Fix bug agent

Duolingo's AI STR Fix runs on the Claude Agent SDK, with Sonnet for most tasks and Opus for harder ones. It does the investigative grunt work engineers used to do by hand, finding relevant code, events, experiments, and recent PRs, writes it to files and the ticket, then proposes a fix. It has fixed release-blocking bugs.

Original writeup

blog.duolingo.com

Duolingo Blog · blog.duolingo.com

We tried “Claude, fix this bug”—here’s what actually worked

Learn how Duolingo built AI STR Fix with Claude to automate bug triage, gather context, reproduce issues, and help engineers fix bugs faster.

Read the writeup

Shelby WengreenPublished Nov 4, 2026Last verified Nov 13, 2026

Claude Agent SDK

Claude Sonnet (most tasks)Claude Opus (harder tasks)

Use case: Investigating and fixing Shake-to-Report bug tickets: gathering context (relevant code, tracking events, experiment states, recent PRs, repro steps) and proposing a fix for engineers to verify
Category: Coding agent
Team size: Built by Duolingo's DevX AI team
Orchestrator / framework: Each task gets a system prompt, a user prompt, skills, and a working directory with the Jira ticket's information and a copy of the relevant repository. Steps gather 'AI context' (relevant code, pre-filtered tracking events, the reporter's experiment states, recent PRs touching the relevant files, repro steps), post it to the ticket, then implement a fix.
Memory: Each step's findings are written as files in the working directory, not the prompt, so later steps can use them.
Evals: not stated
Observability: not stated
Cost (stated in writeup): Not stated
Cost per output (stated): Not stated in the writeup
WhatWeRun estimate: WhatWeRun estimate
Low confidence
$0.50–$6.80 per bug ticket
List prices as of Nov 2026 · estimated Nov 13, 2026 · list prices, no enterprise discounts · not a writeup-stated cost
How we estimated
Stated in the writeup
“We built AI STR Fix using the Claude Agent SDK, with Sonnet handling most tasks and Opus taking on harder ones.” source ↗
“To avoid spending tokens on clearly irrelevant events, we apply a heuristic pre-filter” source ↗
List prices used (Nov 2026)
Claude Sonnet 5.5 (input / output): $2 / $10 per million tokens docs.claude.com ↗
Claude Opus 5.5 (input / output): $4 / $20 per million tokens docs.claude.com ↗
Our assumptions (not from the writeup)
Assumed:Model versions aren't stated; priced at the current Sonnet 5.5 and Opus 5.5 list prices.
Assumed:Low: a typical Sonnet ticket at about 200,000 input and 10,000 output tokens.
Assumed:High: a hard Opus ticket at about 1.5M input and 40,000 output tokens.
Assumed:No prompt-cache discount; excludes infrastructure.
Arithmetic
Low: 200k x $2/M = $0.40 + 10k x $10/M = $0.10 -> $0.50
High: 1.5M x $4/M = $6.00 + 40k x $20/M = $0.80 -> $6.80
What would tighten this
Tokens per ticket and the share of tickets that go to Opus.
Models: Claude Sonnet (most tasks)Claude Opus (harder tasks)
Observability tools: not stated
Other tools: JiraHeuristic pre-filters for tracking events and A/B experiments
Source: https://blog.duolingo.com/ai-bug-fixing-agent/
Shelby Wengreen · Published Nov 4, 2026 · Last verified Nov 13, 2026

corezoid.com

Coding agent

Nov 11, 2026

Corezoid's ticket-routing coding agent

A coding agent triggered by Jira labels that fixes tickets and opens merge requests, with a digital twin of the engineering org deciding which repository, process, or service each ticket belongs to

Stated: ~$2.50 per code change+1

Simulator.Company digital twinLiteLLM gateway+8

aboutamazon.com

Coding agent

Nov 11, 2026

The agent teams that build AWS Emulator

Autonomous agents that write, test, and fix every line of AWS Emulator, a downloadable local emulator of AWS services, with no human-written or human-reviewed code

Kiro (initial prototype)Three agent teams+4