Agentic Engineering sets out the principles. This companion article covers the tools and habits I use to apply them today. It will change as the harnesses and models improve.
When work is too uncertain for one session, I sometimes use a Wayfinder map. It is a parent Backlog.md task with child investigations that resolve important decisions one at a time. I use it selectively. Once the route is clear, conventional implementation tickets are enough.
For larger efforts, Backlog tasks keep the record of scope, dependencies, decisions, and progress across sessions. They replace a single monolithic plan. Tickets coordinate the work, and each ticket's plan should be no more detailed than the work requires.
Before I start, I make important choices explicit and inspect the repository before asking questions it can answer. Pi's sandbox generally confines writes to the current project or a private session workspace, with limited exceptions for my plugin checkouts. Reads outside that scope may require approval; its Bash protections are lexical, not an OS-enforced sandbox. That keeps assumptions visible and the agent's operating context narrow.
I also prefer one focused task per session. When the context grows too large, Pi's /ctx-monitor shows what is consuming it and /handoff carries the decisions and state into a fresh session.
I now keep large reference reads out of the parent context when the task is fact extraction rather than direct reasoning. A Pi hook blocks an unbounded read of a regular file above 16 KiB and points the parent to a shared skill. The skill dispatches a read-only bulk-reader subagent, currently pinned to a cheaper model, with the file paths and question; it returns cited findings, coverage gaps, and uncertainty in at most 600 words.
This is a redirect, not an automatic hand-off. In a Pi usage audit of 200 sessions started September 13–24, 2026, the hook redirected 102 of 3,664 parent read calls (2.8%). That counts attempted broad reads blocked by the hook, not worker dispatches or all opportunities for useful extraction. Bounded reads of up to 350 lines remain available, as do full reads of governing instructions and native documents; search output is outside the gate too. That preserves direct source inspection for debugging, editing, architecture, safety-critical reasoning, and decisions that need exact evidence. The threshold is based on bytes rather than lines because my measurements found bytes to be a more stable proxy for token cost across prose and code. In a controlled two-file comparison, the route cut parent-context tokens by 95.8–95.9%. Applying parent-model prices to those measured tokens and the worker's reported cost yielded an estimated 80–84% saving in combined parent-and-worker cost, not an observed production saving. That is not a general benchmark: I measure parent-context reduction, combined cost, latency, and extraction accuracy separately. Spotify's roughly 90% figure motivated the pattern, not a result I claim for this setup.
Independent Backlog tasks can run in separate Git worktrees. Dependent work stays sequential. Claude Code supports this through --worktree sessions and worktree-isolated subagents. In Pi, I use pi-worktree-agents for the same approach.
The goal is not to maximise the number of tickets. It is to keep the backlog current and the important reasoning close to the code, where Git makes completed work searchable without creating another system that can drift from the implementation.
My coding harness is a public example of this split. The shared layer contains user-authored resources both harnesses can use: AGENTS.md, skills, and command-safety and read-routing regression contracts. Pi keeps its own extensions, prompts, and agents. Claude Code keeps its own hooks, settings, themes, and agents. Credentials, sessions, caches, provider configuration, and generated state remain local.
A link script installs the shared resources into both Pi and Claude Code without maintaining copies. Its optional package step installs the Pi packages listed in the version-controlled manifest. These are currently pi-worktree-agents and pi-status-footer. Local Pi settings record what is installed on a given machine. I added and removed the artifacts in response to observed failure modes. Keeping the shared layer small means another harness can join later without moving its source of truth.