SAGE agents do the cheap prep work. Claude Code does the expensive execution. Every Opus token goes toward building, not orienting.
Claude Code (Opus 4.6, 1M context) can hold entire codebases and execute complex multi-file changes in one shot. But without pre-structured work, sessions spend 30-50% of their expensive context just figuring out what to do.
Route each activity to the cheapest model that can handle it. SAGE agents run continuously on Sonnet/Haiku. Claude Code fires on-demand for complex execution.
The handoff mechanism is a self-contained Execution Prompt — a structured, ready-to-paste instruction block placed in a Handoff Queue. Claude Code picks it up and executes immediately. No research phase needed. 90-100% of context goes toward building.
Each SAGE agent owns a slice of the prep work. Together they produce execution-ready work packages that Claude Code can act on without orientation.
Scans trackers, identifies highest-priority work, gathers all context (files, deps, Figma nodes), decomposes tasks into atomic steps, writes structured Execution Prompts.
Audits session logs, backlog items, Notion organization. Flags stale items, duplicates, and orphan pages. Produces actionable remediation lists.
Checks deployment health across all services. Flags incidents with severity and recommended action. Produces incident reports that escalate to execution prompts when code fixes are needed.
Reviews specialist outputs, resolves priority conflicts across projects, approves escalations to Claude Code, and produces weekly operational summaries.
Every task flowing from SAGE to Claude Code must be a self-contained instruction block. No additional research needed.
## SAGE Execution Prompt — [SAG-XX] ### Objective [One sentence: what needs to happen] ### Context - Project: [Atlas / OpenClaw / Delphi OS / KAA] - Lane: [PORTAL / BACKEND / CODEX / OPENCLAW / INFRA] - Source: [Build Tracker BO XX / Backlog ID / Incident] - Priority: [Critical / High / Medium] - Dependencies: [What must be true before starting] ### Files of Interest - path/to/file1.ts — [why this file matters] - path/to/file2.tsx — [what needs to change] ### Execution Steps 1. [Specific, imperative step] 2. [Next step] 3. [...] ### Acceptance Criteria - [ ] [Measurable outcome] - [ ] next build passes with 0 errors ### References - Figma: [node ID if UI work] - Related: [SAG-XX, Backlog ID]
A continuous pipeline where SAGE agents prep and Claude Code executes — no gaps, no wasted context.
Project Manager structures the work, gathers all context, writes a self-contained Execution Prompt with files, steps, and acceptance criteria.
Execution Prompt placed in the Handoff Queue with Status: Pending, Direction: SAGE to CC, priority level, and target session lane.
Session Start auto-behavior checks the Handoff Queue. Finds Pending packet, claims it, reads the Execution Prompt.
No research phase needed. 90-100% of context goes toward building. Multi-file changes, architecture decisions, full build sessions.
Post-build verification, session log with resume point, tracker updates. SAGE agents pick up completed status and prep the next task.
The same session budget, dramatically more productive. SAGE pre-digests everything so Opus only does what Opus is for.
2-3x more productive Opus sessions at the same token cost. Every dollar spent on SAGE prep ($0.25-$0.80) saves $2-8 of Opus context that would have been wasted on orientation. The ROI compounds with every session.