Designing Custom Agentic Systems

Resources

• Design the working session deliberately — Claude Code’s best-practices guide connects direct instructions, file references, images, exploration, planning, and verification. Use it to decide when inline context and a focused session are sufficient before introducing more agents or automation. https://code.claude.com/docs/en/best-practices

• Skills as executable prompt expanders — Claude Code documents reusable skills with arguments, supporting files, dynamic context injection, invocation controls, and forked execution. This turns repeated guidance into a maintained interface while preserving the distinction between automatic selection and explicitly invoked workflows. https://code.claude.com/docs/en/skills

• A metaprompt that produces a starting draft — Anthropic’s cookbook notebook generates a prompt template from a task and optional variables, then provides a way to test it. Its explicit single-turn scope and non-optimality caveat are useful for teaching metaprompts as drafting aids, not substitutes for custom planning and evaluation. https://github.com/anthropics/claude-cookbooks/blob/main/misc/metaprompt.ipynb

• Predefined specialists in a coding harness — Claude Code’s subagent documentation covers file-based definitions, descriptions used for delegation, tool restrictions, models, memory, and execution settings. Use it to design clear responsibilities and automatic triggers rather than multiplying agents with overlapping instructions. https://code.claude.com/docs/en/sub-agents

• Programmatic and dynamic delegation — The Claude Agent SDK guide explains programmatically supplied subagents, context inheritance, parallel invocation, and resuming delegated work. It provides the bridge from a fixed roster to task-specific agent definitions and explicit orchestration. https://code.claude.com/docs/en/agent-sdk/subagents

• Harnesses for long-running work — Anthropic describes initializer and coding-agent roles, progress files, incremental work, and verification across context windows. It is a concrete reference for designing continuity and completion checks rather than asking an agent to keep working indefinitely. https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

• Worktrees isolate files, not responsibilities — Git’s worktree manual explains multiple working trees, branch constraints, locking, and cleanup. Use it to give parallel coding sessions separate execution spaces while making task ownership, integration, and merge review explicit system-design decisions. https://git-scm.com/docs/git-worktree

• Hooks turn policy into event-driven behavior — Claude Code’s hook reference specifies lifecycle events, input and output contracts, exit codes, asynchronous hooks, and tool-related interception. It supplies implementation details for injecting context, validating actions, and feeding results back into a custom workflow. https://code.claude.com/docs/en/hooks

• Session polling versus durable scheduling — Claude Code documents repeated prompts and one-time reminders alongside the lifecycle limits of session-scoped tasks. Use it to choose deliberately between a temporary status-check loop and automation that needs to survive independently of an open session. https://code.claude.com/docs/en/scheduled-tasks

• Heartbeats as bounded check-ins — OpenClaw documents heartbeat cadence, active hours, notification behavior, context choices, and operating costs. It is a concrete example of proactive agent activation, useful for distinguishing a periodic judgment call from an exact scheduled job or an event handler. https://docs.openclaw.ai/gateway/heartbeat

• Monitoring the custom harness — Claude Code’s monitoring guide explains its OpenTelemetry integration and usage telemetry. It gives the lesson an observation layer for seeing whether background and team workflows consume resources or encounter failures, rather than equating silence with success. https://code.claude.com/docs/en/monitoring-usage

• Shared instructions and agent-maintained memory — Claude Code distinguishes maintained project instructions from automatically accumulated memory and explains their loading and scope. Use it to separate team-owned conventions from local learnings and to reason about stale or inappropriate context reaching future tasks. https://code.claude.com/docs/en/memory

• Settings need ownership and precedence — Claude Code documents user, project, local, and managed configuration scopes and how conflicting values resolve. This is the control-plane reference for sharing tool and permission settings without silently overriding individual or organizational boundaries. https://code.claude.com/docs/en/settings

• Repository events become an agent pipeline — Claude Code’s GitHub Actions documentation shows agents responding to repository activity within CI workflows. Use it to connect shared instructions and command-line execution to bounded triggers, repository permissions, reviewable changes, and automated feedback. https://code.claude.com/docs/en/github-actions