LEAN AGENTIC CODING: ADD A MEMORY LAYER AND MAKE SKILLS PORTABLE
Practitioners are converging on lean, memory‑equipped agents and cross‑platform skills as the practical way to use AI for coding. A hands‑on guide argues codin...
Practitioners are converging on lean, memory‑equipped agents and cross‑platform skills as the practical way to use AI for coding.
A hands‑on guide argues coding assistants need persistent context beyond the chat window so they stop asking the same questions and defaulting to the wrong stack; think per‑repo rules files and long‑term memory services, not just bigger prompts Towards Data Science.
Two field reports back this up: one distills what worked and what didn’t when “agentic” coding met real tasks DEV, and another shows an agent that searches GitHub first, then emits platform‑specific skills (Cursor .mdc, Manus/IBM Bob YAML/JSON) using Claude’s Managed Agents beta to kill cross‑tool rewrite tax Medium.
A build guide pushes the same theme: use AG2 (formerly AutoGen) with tight tool use and MCP, avoid multi‑agent swarms, and evaluate on reliability, latency, and cost—ship one agent with a clear goal, then expand Substack.
Waste from repeated context priming and tool translation adds up; memory layers and portable skills cut cycle time and cost.
Smaller, evaluable agents are easier to operate in production than flashy multi‑agent swarms.
-
terminal
Add a per‑repo rules/memory file to your assistant and measure change in tokens per task, back‑and‑forth corrections, and task time over two sprints.
-
terminal
Build one AG2 agent with an MCP tool for a narrow SRE/ops task (e.g., log query + rollback) and compare reliability/latency/cost to today’s runbook.
Legacy codebase integration strategies...
- 01.
Introduce long‑term memory via rules files that encode stack, ports, conventions, and guardrails; version them with the repo.
- 02.
Wrap existing CLIs/services as MCP tools to keep data locality and audit trails while giving the agent structured capabilities.
Fresh architecture paradigms...
- 01.
Standardize an internal, neutral skill schema and auto‑generate platform‑specific formats (.mdc, YAML/JSON) per target tool.
- 02.
Start with one goal‑driven agent, one planner model, and 2–3 vetted tools; add complexity only after proven wins.