Keep agent sessions small

Anthropic's advice for getting more from Claude Code sessions comes down to one rule: keep each context small and relevant.

Start a fresh conversation for each task. Pick the model and effort before work begins. Point Claude at the exact files, keep command output quiet, move noisy work into another context, and compact a long-running task before its prompt cache goes cold.

Saggar is built around the same rhythm. A new task gets a named terminal, with its model chosen at launch. Independent work can move into another agent terminal or worktree. Loops, test watchers, servers, and log tails live in monitors, while short commands run in Quick Runs, so their output doesn't pollute an agent's conversation.

Those separate sessions stay manageable because Saggar tracks which are Needs you, Working, Idle, Finished, or Failed. You can leave working sessions alone, answer the queue when one blocks, and resume a supported agent's exact conversation later.

Saggar can also choose a quiet moment for compaction. Turn on Compact idle Claude chats, and after you have been away for 30 minutes Saggar will compact an idle Claude chat whose context is at least half full. It never interrupts a working or waiting chat, and records confirmed compactions in the summary you see when you return.

Claude Code's /compact and /autocompact commands remain useful when you want direct control. For a different task, a fresh Saggar terminal is still the cleaner boundary.