Oct Hands-On Agent Building

A single Agent with on-demand loaded Skills is often faster and more stable than multi-Agent handoffs, loses less context, and also reduces latency.

When there are many tasks, you can schedule in layers: one master coordinator oversees everything, several project managers divide up the tasks, and a batch of execution Agents run automatically overnight.

Agent memory shouldn't just be a pile of Markdown files. With more than 10,000 files, switch to SQLite queries, which use about 40 times fewer tokens to answer the same question.

Retrieval that mixes vectors with BM25 keyword matching, followed by reranking, prevents semantic drift from ranking irrelevant content first.

Set a round limit for Agents reviewing each other's code: after the third round, have the author rethink the approach, then decide how many reviewers are needed based on the risk of the change.

MCP can plug specialized abilities into Agents: Serena locates code by function to save tokens, and an Unreal plugin lets Claude Code operate the editor.

Connect Codex and Claude Code as sub-bots under a master Bot, dispatching by task type and consuming each account's own quota.

Only by hand-writing a rough Agent loop, such as a 200-line version with four tools, can you understand how tool results are fed back and how context is compressed.