Get more from your AI base plan with portable project state.
Orchestration is not running the most expensive model on every prompt. It is giving each model one clear job, preserving the result outside the chat, and moving verified state forward when the next phase needs a different tool.
The base-plan orchestration loop
Use a strong reasoning model to define the goal, constraints, acceptance tests, and task boundaries. Save the decisions outside the chat.
Send narrow edits, searches, formatting, and test execution to cheaper or local models. Reserve frontier usage for architecture, ambiguity, and difficult debugging.
Run tests and inspect the diff before the next phase. Council can challenge risky conclusions; the handoff records what actually passed.
Before context pressure or usage exhaustion, produce a lean continuation packet and start the next phase in clean context.
Where base-plan usage gets wasted
Most waste is not the final code generation. It is repo discovery repeated by every new agent, long transcripts full of abandoned branches, expensive models doing mechanical work, and fresh sessions repeating failures because the previous state was not durable.
What ShardStitch orchestrates
Compresses noisy context while pinning decisions, constraints, and rejected approaches that must survive.
Uses multiple specialist perspectives when a decision is risky enough to justify extra inference.
Keeps project decisions, conventions, bugs, and verification state outside any one provider chat.
Shapes the current goal, disk truth, failed paths, dependency risk, and next action for the next AI tool.
A practical model budget
Use the strongest model for: requirements, architecture, hard tradeoffs, unfamiliar failures, security-sensitive work, and final review.
Use cheaper or local models for: focused searches, small edits, test generation, documentation, repetitive migrations, and isolated refactors with clear acceptance criteria.
Start a fresh session for: implementation after planning, verification after implementation, a deep debugging branch, or any task after the current chat accumulates stale assumptions.
Failure-safe orchestration
Normal orchestration assumes every agent finishes and writes the note. Real workflows hit rate limits, crashes, stale context, deleted chats, and partial edits. ShardStitch reads the state that survived, separates verified evidence from inference, and gives the next agent a safe restart point.
Related workflows
FAQ
Can orchestration increase my AI plan limit?
No. It cannot change provider quota. It helps you spend less of that quota on repeated discovery, stale context, and duplicated work.
Should every task use the strongest model?
No. Use frontier reasoning where mistakes are expensive. Route bounded mechanical work to cheaper or local models.
What happens if an agent dies before writing the handoff?
ShardStitch reconstructs a continuation packet from surviving project and session evidence, then marks verified facts separately from inferred next steps.