Agentic AI Coding Workflow — Multi-Agent with Stacked Diffs
AI · · 1 min read · x.com ↗
Agentic AI Coding Workflow — Multi-Agent Pattern
Author: Jacob Bartlett (@jacobtechtavern)
Source: X/Twitter
Date: 2026-05-13
Original Content
The developer workflow:
- Plan is in place with stacks and per-stack verification steps
- Codex codes for hours while developer works on other tasks (multiple separate repo checkouts)
- Result is a testable, mostly-there feature in stacked diffs
- Code review in chosen client with inline comments across the stack
- Tell the agent: "Go through the PR comments, one stacked diff at a time, addressing the issues flagged. Make changes on the appropriate point in the stack, and reply to each PR comment to confirm."
Key insight: The agent correctly oriented itself across a complex multi-hour stack of diffs and made changes at the appropriate points. Codex is cited as the only agent trusted for this.
Multi-agent breakthrough pattern: 1 Writer + N Readers on the same checkout
- Reader #1: Fetches PR comments, gathers context, proposes fixes
- Reader #2: Reviews code in each stack independently, outputs improvements/simplifications
- Writer: Reads the markdown outputs and implements changes
- Optional idle agent at top of PR stack for manual testing (logs, test data, output review)
Practical limits: 1 Writer + 2 Readers is the mental processing limit.
Full article: https://blog.jacobstechtavern.com/p/ios-at-an-ai-unicorn
Summary
Detailed agentic AI workflow that combines:
- Stacked diffs for clean, reviewable PR structure
- Multi-agent parallelization (1 Writer + 2-3 Readers)
- Asynchronous work (developer does other tasks while agent codes)
- PR comment-driven iteration (agent addresses review comments automatically)
- Separate repo checkouts to avoid agent conflicts
Metrics
- Views: 15
- Full Article: https://blog.jacobstechtavern.com/p/ios-at-an-ai-unicorn
Relevance to FACorreia Projects
HIGH RELEVANCE — directly applicable to Norviq, Loci, and ObsidianBrain iOS development workflow. The multi-agent pattern (Writer + Readers) overlaps with the Hermes multi-Agent skill architecture (CTO, PM, Dev, QA, Ops). Key takeaways:
- Stacked diffs enable granular code review
- Multi-agent parallelization requires careful coordination to avoid conflicts
- 1 Writer + 2 Readers is the practical limit for effective coordination
- Codex is trusted for complex, multi-hour multi-file changes; other agents less reliable
This is an iOS developer working at scale with AI agents — the exact workflow pattern you should consider for your iOS projects.