> For the complete documentation index, see [llms.txt](https://breakpoint-journal.gitbook.io/breakpoint/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://breakpoint-journal.gitbook.io/breakpoint/ai/ai-tooling-for-swe-tricks-workflows-and-what-worked-at-coinbase.md). # AI Tooling for SWE: Tricks, Workflows, and What Worked at Coinbase In this overview, I'll share some of my personal notes from my time as a full-time SWE at Coinbase (from both Base and Institutional organizations) ## **General Tips** * Create or modify [agents.md](http://agents.md/) * Use MD files to transfer context with chats * Chat’s pay particular attention to the beginning and end of the chats, so any chats that are at 70% or more context capacity should be discarded * System Prompts > [Agents.md](http://agents.md/) > Context + Rules > User Commands * Plan > Execute > Review > Feedback > Plan > … ## **Context** **Agents vs Sub-Agents**\ An agent is just a session. **Different Agents**\ from Alex's talk, using different models to do different tasks (i didn't know about gpt-5.2 and how its better for non-coding related tasks) ## **Sub-Agents** * * * * ## **Secure Coding Agent Container (SCAC)** **s** ```shell Usage: cd your-project-folder Usage: scac [agent-args...] Available commands: claude Claude Code CLI (bind-mount mode) claude-copy Claude Code CLI (copy-based mode) amp AmpCode CLI (bind-mount mode) amp-copy AmpCode CLI (copy-based mode) exec Execute command in running container pipe Copy file from host to running container sync-back Sync container changes back to current directory (incremental mode) sync-back auto Automatically sync changes from all running SCAC containers ps List running SCAC containers cleanup Remove stopped SCAC containers Copy-mode flags: --detach Start container in detached mode and return container ID Examples: scac claude --print # Run claude --print --dangerously-skip-permissions scac claude-copy --help # Run claude --help --dangerously-skip-permissions in copy mode scac claude-copy --detach # Start detached container, returns container ID scac amp --version # Run amp --version scac amp-copy # Run amp in copy mode scac exec claude --help # Execute claude --help in running container scac exec bash # Open bash shell in running container scac pipe /path/to/file # Copy file from host to running container scac sync-back # Sync container changes back to current directory scac sync-back auto # Automatically sync from all running containers scac ps # List running SCAC containers scac cleanup # Remove stopped SCAC containers ``` ## **OpenCode** ### **MCPs** ### **Plugins** ### **Oh-My-OpenCode** Most agent harnesses give you one agent with tools. Oh My OpenCode gives you a team of specialized agents that work like a real engineering organization. The key insight: LLMs aren't interchangeable. Different models excel at different tasks. GPT 5.2 is a debugging god. Gemini 3 Pro dominates frontend. Claude Sonnet is great for documentation. Oh My OpenCode orchestrates them by purpose, not by availability. #### **Commands** | Command | Description | | ------------------- | ------------------------------------------------------------------------------------ | | `/code-review` | Review code for bugs and CLAUDE.md compliance | | `/debugging` | | | `/snowflake` | | | `/sql-optimization` | | | `/commit` | Create a git commit | | `/commit-push-pr` | Commit, push, and create a PR | | `/clean\_gone` | Clean up gone git branches | | `/review-pr` | Comprehensive PR review using specialized agents | | `/feature-dev` | Guided feature development with architecture focus (this is like a "plan" workflow!) | | `/revise-claude-md` | Improve CLAUDE.md files | | `/create-plugin` | Create a new Claude Code plugin | | `/new-sdk-app` | Create a new SDK app | | `/ralph-loop` | Run Claude in autonomous loop mode | | `/cancel-ralph` | Stop ralph-loop | | `/hookify` | Configure hooks | | `/configure` | Configuration helper | | `/list` | List available items | | `/help` | Get help | | `ultrawork (ulw)` | Maximum effort, great for complex features (use sparingly) | * Normal prompts → Opus 4.5, solid experience. * Add plan/analyze → Deeper multi-agent reasoning. #### **Agents** * Sisyphus (Opus 4.5 High): The orchestrator—thinks like a tech lead, delegates like a manager * Hephaestus (GPT 5.2 Codex): Autonomous deep worker for goal-oriented, end-to-end execution * Oracle (GPT 5.2): Architecture, debugging, strategic consultation * Librarian (Sonnet 4.5): Documentation, OSS implementations, cross-repo analysis * Explore (Haiku 4.5): Fast, cheap contextual grep—fires in parallel * Frontend Engineer (Gemini 3 Pro): UI/UX specialist #### **Installation** [Opencode + OmO: 10x your cbcode experience](https://docs.google.com/document/d/1V-OStW2ipM1N2BElIBPZJF38Ex7T5xrEuRC0tUT9-aI/edit?tab=t.0) #### **Documentation** ## **WorkTrunk** See the WorkTrunk section of the following [guide](https://docs.google.com/document/d/18BRUJfjaB1JndfA7L2TuRqqeRvhcP5R6gkMo64r_2ww/edit?tab=t.0) for more information. ### **Features** * [ ] Add post-start hooks ## **Vibe Kanban** ### **Features** * Add PR feedback directly in UI * Create PRs * Custom Prompt Tags: save prompts as tags that you can include with certain tasks ### **Issues** #### **Anthropic API Key Missing** Copied ANTHROPIC\_AUTH\_TOKEN from \~/.claude/settings.json into \~/.config/opencode/opencode.json. ## **GSD (Get-Shit-Done)** \ The concept behind GSD is that it utilizes the scientific method (i.e. plan, hypothesize, execute, verify, rinse-repeat) to complete features. It creates a very structured file system and documents phases of your project in .md files before executing. **Highly recommended to execute with the following otherwise you’ll be approving 50x before the project can even finish:** ```shell claude --dangerously-skip-permissions ``` To see an example of how the phases work, see the [README.md](http://readme.md/) ## **Ralph Loops** [How to Ralph](https://docs.google.com/document/d/1F7x2C7gLQtKxWZLhWTOA8CraxXAE9cKMWDAB4wjrkA8/edit?tab=t.0#heading=h.tb86d3xooru7) ```shell 🚀 Usage: 1. Run 'cbcode' in this directory 2. Use /prd or ask Claude to 'create a prd' to generate a PRD 3. Run ./ralph.sh to start the autonomous implementation loop 🛡️ Security mode: 🔒 Secure containerized execution (default) 💡 Ralph execution modes: ./ralph.sh # Default: SCAC copy mode (isolated in a container) ./ralph.sh --bind # SCAC bind-mount mode (your file changes are visible) ./ralph.sh --no-scac # Direct execution (not protected by SCAC) ./ralph.sh 100 # Run a ralph loop with 100 iterations ./ralph.sh --help # Show all options ``` If you’re ever unsure how many loops to run for, you can ask Claude what is suggested given the PRD. For example, if you have **three** phases in your PRD, you could set Ralph to run for **six** iterations such that each phase can get a second review. I also generally run with **`–bind`** s.t. I can see the output of the Ralph loop as it works. ### **Ralph Orchestrator** \ An improved version of the Ralph-Wiggum method that also adds a dashboard running localhost for tracking jobs. ### **Willie** A knowledge indexer that utilizes Ralph loops to generate a Github repository of team-specific information. ```shell # Basic run (interactive mode, 50 iterations) ./willie # Non-interactive with custom iterations ./willie -y -n 10 # Dry run (shows what it would do without executing) ./willie --dry-run # Full help ./willie --help # Reset collected data ./willie clean # Clears data, keeps team context ./willie clean --all # Full reset including team context ``` The concept is you can give it a set of tasks, human readable instructions, explaining where to fetch data from. I’m curious why it’s useful to have this in a single repository however. ## **N8N** ### **Base-Account On-Call** See the Base Account [flow](https://coinbase.slack.com/archives/C0ABND53KHA/p1769633170824189) for an example of how n8n integrates with PagerDuty, DataDog, Google Sheets, and Linear to automatically create backlog tickets to improve monitor SNR (sound-to-noise ratio). It’ll propose tickets that add subtags to monitors or update insufficiently documented runbooks etc. ## **Project Ideas** ### **Hacker** * Auto-reviews online CVEs and security issues * Reviews over code from a security perspective * Automatically runs updates ### **FixerUpper / XFit** Automatically run overnight and throughout the day finding tasks to cleanup and optimize the codebase. Auto integrates with slack to produce a round-up. I need to go through each one and approve which end up on Slack however. * Monitor DD for error logs and pages, setup investigator * Checks for vulnerabilities and issues * Check repository’s for out-of-date documentation ## **Action Items** * [x] ~~Create a Secure Code Agent container and let it use your CLI tools~~ [~~https://github.cbhq.net/wallet/secure-code-agent-container~~](https://github.cbhq.net/wallet/secure-code-agent-container) * [x] ~~Visual orchestrator for agents:~~ [~~https://www.vibekanban.com/~~](https://www.vibekanban.com/) * [x] [~~Watch Brunos’ on-call AI~~](https://coinbase.slack.com/archives/C0ABND53KHA/p1769633170824189) * [ ] Ambient Agents * [x] ~~GSD (Get-Shit-DOne)~~ * [ ] [Graham Perich - AI Retool 2/26 - Agentic Software Engineering Refresh](https://docs.google.com/document/d/1YCuU_Bmo81hENijcd4XOrYhR-LrppTNGTTAb7M6UnBw/edit?tab=t.0#heading=h.mhn2mmdb5efe) * [ ] Create user commands * [ ] Modify my [agents.md](http://agents.md/) * [ ] [Create SubAgents](https://code.claude.com/docs/en/sub-agents) (act like personalities)\ \- \[ ] Tell one agent to be extremely skeptical in Plan mode and don’t make any questions without consulting me. Ask me lots of questions, etc.\ \- \[ ] Break down PB\&J sandwich problems ## Slides ![](https://content.gitbook.com/content/mGhYQRY1OEeL4zywZ9zS/blobs/08XyBMm8zVHy3qHjBbdn/Unknown%20image)\ ![](https://content.gitbook.com/content/mGhYQRY1OEeL4zywZ9zS/blobs/REmW26eJgJCOeeWZ2O1a/Unknown%20image)