# What's your agentic workflow? — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/ae458e664d26e2df

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-11-10 18:28 ingested 2026-11-11 02:37

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xsuln3/whats_your_agentic_workflow/)

## retained text



```
I'm curious how other people work with coding agents day to day. Single session and manual review, subagents, some own harness?

Mine for bigger work is workflow template for Claude Code. Diagram attached, top is how I worked before, bottom is how it works now.

Before it was prompt, agent builds, I test by hand or let agent click clack, "fix this", 3 or 4 rounds, ship, some bugs in prod anyway.

Now:  
\- I write project brief with goal, acceptance criteria per stage and non-goals  
\- stages run in order, 3 tasks in parallel  
\- every task has builder, checker and curator agent  
\- checker tests output against criteria and can open follow-up task when it finds gap (max 4 per task)  
\- curator writes lessons to knowledge base, next tasks and next runs read it  
\- at the end one e2e run, report, editor agent reviews report  
\- limits: spend cap, max 24 tasks, stage which fails review 2 times stops loop and goes to report

For small fixes I still use normal single session, harness is too slow and uses too many tokens for that.

Thing I still don't like is hard caps on checker. Without them it creates work forever, with them sometimes it cuts real fix.

So what does your setup look like? And if you let agent check agent, how do you limit it?
```



document fetched 2026-11-11 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xsuln3/whats_your_agentic_workflow
