AGENTIC CODING: MAKE THE TODO/PLAN A FIRST-CLASS TOOL
A popular HN breakdown shows you can recreate a Claude Code–style assistant in a few hundred lines by centering the loop on a dynamic TODO/plan and persisting i...
A popular HN breakdown shows you can recreate a Claude Code–style assistant in a few hundred lines by centering the loop on a dynamic TODO/plan and persisting it with explicit tools instead of relying solely on the model’s context. Practitioners note large performance drops when planning/TODO is disabled and warn that production-grade context and memory management is the real work; YouTube summaries also claim Claude Code 2.1 adds sub-agents/skills and LSP hooks, but verify against official docs.
Explicit planning state improves reliability, observability, and repeatability versus context-only prompting.
This pattern is fast to prototype but surfaces the true cost of context, memory, and guardrails you’ll need in production.
-
terminal
Prototype a repo assistant that persists a structured TODO list (timestamps, branches, outcomes) to disk/Redis and compare task success with and without planning enabled.
-
terminal
Evaluate LSP-based local edits versus patch generation on a target service and measure diff quality, latency, and reviewer acceptance.
Legacy codebase integration strategies...
- 01.
Start read-only in monorepos and gate writes through CI; keep agent plan state outside the repo to avoid noisy churn.
- 02.
Map agent actions to code ownership and ticketing rules to prevent cross-boundary changes and ease rollback.
Fresh architecture paradigms...
- 01.
Define a plan/TODO schema and tool interface on day one, persist it in a durable KV, and log every step for traceability.
- 02.
Standardize editor/LSP setup and patch/rollback mechanics early to keep agent loops deterministic.