Working with AI Harnesses

Resources

• Claude Code across working surfaces — The official overview maps terminal, IDE, desktop, web, and integration entry points. Use it to identify what a coding harness supplies beyond the model and what changes when the same work moves between synchronous chat and delegated execution. https://code.claude.com/docs/en/overview

• Codex as a desktop coordination surface — The Codex app documentation introduces a workspace for agent threads and project-based work. It supports comparing desktop supervision and parallel work with a single conversational interface rather than comparing models by brand alone. https://learn.chatgpt.com/docs/app

• Codex in the terminal — The CLI documentation provides a contrasting interface for repository-local agent work. Use it to examine how environment access, approvals, and terminal interaction affect an otherwise similar context–act–verify loop. https://learn.chatgpt.com/docs/codex/cli

• A harness inside your own program — The Codex SDK documents controlling agent threads from application code. It gives teachers a concrete bridge from interactive use to custom scripts, repeatable runs, and dashboards that supervise work instead of merely displaying chat output. https://learn.chatgpt.com/docs/codex-sdk

• Cursor rules as scoped personalization — Cursor’s rules documentation explains reusable instructions and how they attach to project work. Use it to separate portable conventions from harness-specific activation behavior when translating a workflow between IDE and CLI environments. https://cursor.com/docs/rules

• Pi’s minimal, extensible harness — Pi’s official coding-agent README describes provider support, model selection, sessions, extensions, and programmatic interfaces. It is a useful counterpoint to larger integrated products: inspect which workflow capabilities belong to the core and which must be deliberately supplied through extensions or scripts. https://github.com/earendil-works/pi/tree/main/packages/coding-agent

• AGENTS.md as a portable repository contract — This project documents a shared Markdown convention for repository instructions such as setup, tests, and coding practices. Use it as a harness-agnostic starting point, while checking each harness’s discovery and precedence behavior instead of assuming identical interpretation. https://agents.md/

• Worktrees separate files, not responsibility — Git’s worktree reference explains multiple working trees attached to one repository. It grounds multi-threaded agent work in a real isolation mechanism and opens discussion of integration checks, shared repository state, and who owns merging competing changes. https://git-scm.com/docs/git-worktree

• Hooks turn conventions into feedback — Claude Code’s hooks reference documents lifecycle events and programmable responses around agent activity. Use it to distinguish contextual reminders from executable checks that load context, validate changes, or block actions at defined points. https://code.claude.com/docs/en/hooks

• Issues and pull requests as agent entry points — Claude Code’s GitHub Actions integration connects repository events and mentions to agent work. It provides a concrete example of issue enrichment, proposed fixes, and communication pipelines whose permissions and verification must remain explicit. https://code.claude.com/docs/en/github-actions

• Reviewing a changing patch series — Git’s range-diff compares two versions of a series of commits. It gives the lesson a precise tool for layered review: inspect what changed after feedback without confusing the latest patch series with an undifferentiated final diff. https://git-scm.com/docs/git-range-diff

• An issue system as a programmable interface — Linear’s GraphQL guide documents querying and mutating issue-management data. Use it to design connector-backed issue enrichment and dynamic progress dashboards, separating read-only reporting from writes that alter team commitments or notify people. https://linear.app/developers/graphql

• Multi-agent code review as a harness feature — Claude Code’s review documentation describes an agent-based pull-request review workflow. Use it to discuss enriched findings and review orchestration while preserving the distinction between an agent’s assessment, deterministic verification, and final human approval. https://code.claude.com/docs/en/code-review