AI + Platform engineers at scale often... Playbook

Platform engineers at scale often juggle multiple concurrent technical workstreams (test suite migrations, dependency upgrades, architectural refactoring) while holding complex context across several git worktrees simultaneously. Without structured support, this context-switching creates cognitive overhead and burnout risk. An effective AI workflow allows engineers to carry three independent projects in parallel at the pace a single project would normally require, using LLMs to maintain each worktream's momentum independently without losing technical context.

Start each workstream with a planning session using Prometheus (the planner agent in Oh My OpenAgent) to iron out implementation details and generate a plan.md file. The plan.md includes a high-level overview with context, followed by specific tasks and orchestration details. Review the high-level overview and context section to validate that Prometheus understood your intent correctly. If the statements don't match your intent, either rewrite the plan.md directly or re-prompt Prometheus to adjust until satisfied. Once the high-level direction is validated, dispatch Sisyphus (the builder agent) against the relevant git worktree to execute the plan while you move to the next planning cycle. Overlap planning and building: let one agent build while you plan the next workstream, then review the first agent's output while the second builds. Keep each workstream on a separate issue to maintain clear context boundaries and prevent confusion across parallel work. Prioritize different types of work in separate cycles (dependency updates, test suite migrations, tooling development) rather than parallelizing similar tasks, which creates context collision.

Dramatic acceleration in issue throughput: completing three or more issues in the time previously required to start a single issue. The workflow unlocks entire categories of work that were previously impossible to justify — maintenance and refactoring tasks that improve system health but don't directly deliver new features. These can now run as parallel workstreams alongside primary feature development, enabling both sustained platform improvements and customer-facing value delivery simultaneously, rather than forcing a choice between them.

Do not attempt to parallelize multiple workstreams within the same technical domain simultaneously. Trying to overlap planning, building, and review cycles across similar work (e.g., multiple test migration tasks) creates cognitive collision and makes it difficult to keep implementation details straight. Instead, run one OMO session per domain and allow sequential domain work to proceed. Consider offloading review guides, architectural documentation, and comms to Claude Desktop for editorial polish rather than managing everything through the OpenCode TUI. Anchor high-level direction and context in persistent external systems (Jira, kanban boards) rather than relying solely on memory or TUI state to maintain clarity across workstreams.

Memory/cross-session state. AGENTS.md files are a good starter, but there's nuance that's hard to get right. Exploring options, but nothing has quite solved.