Quality assurance · Production

How Delivery Hero's agent merges 100+ pull requests a day with Claude

The problem

Delivery Hero's engineering teams across dozens of regional operations each had different build systems, infrastructure setups, and engineering cultures inherited from independent acquisitions. Centralizing AI tooling was hampered by procurement delays and a decentralized experimentation phase that made consistent measurement and scaled adoption impossible.

Workflow diagram · grounded in source

Task assigned via Jira

trigger

“Herogen picks up tasks assigned through Jira”

Herogen writes and tests code

ai_action

“Herogen picks up tasks assigned through Jira, writes the code, runs and iterates on tests, then submits the result as a pull request”

Council of agents reviews code

validation

“A "council of agents" that includes both Claude and Gemini reviews the code from different perspectives before a human does a final check. Using multiple models for review is intentional: it reduces the chance that blind spots in any sin…”

Human final check

human_review

“before a human does a final check”

Pull request submitted

output

“then submits the result as a pull request, a proposed code change for review”

Reported outcome

Herogen now merges over 100 pull requests per day at an 85% success rate, achieving 18 times the original Q1 2026 quarterly goal.
Claude accounts for 95% of LiteLLM model requests across central engineering teams, with a 9-to-2 preference for Claude emerging from a survey of the CTO group.

Reported metrics

Merged pull requests per day>100

Task success rate85%

Q1 2026 quarterly goal achievement18x the original quarterly goal

Claude model usage share95%

Show all 7 reported metrics

merged pull requests per day>100

task success rate85%

Q1 2026 quarterly goal achievement18x the original quarterly goal

Claude model usage share95%

CTO group model preference9-to-2 preference for Claude over the next closest option

cross-repository change turnaroundpreviously a multi-day effort, completed in minutes

Herogen share of all PRsabout 9%

Reported stack

ClaudeClaude Opus 4.5Claude CodeHerogenVertex AILiteLLMJiraGemini

Source

https://www.anthropic.com/customers/delivery-hero

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Herogen now merges over 100 pull requests per day at an 85% success rate, achieving 18 times the original Q1 2026 quarterly goal.

What tools did this team use?

Claude, Claude Opus 4.5, Claude Code, Herogen, Vertex AI, LiteLLM, Jira, Gemini.

What results were reported?

Merged pull requests per day: >100; Task success rate: 85%; Q1 2026 quarterly goal achievement: 18x the original quarterly goal; Claude model usage share: 95% (source-reported, not independently verified).

How is this quality assurance AI workflow structured?

Task assigned via Jira → Herogen writes and tests code → Council of agents reviews code → Human final check → Pull request submitted.

Herogen now merges over 100 pull requests per day at an 85% success rate, achieving 18 times the original Q1 2026 quarterly goal.Claude accounts for 95% of LiteLLM model requests across central engineering teams, with a 9-to-2 preference for Claude emerging from a survey of the CTO group.