AI coding agents often fail before they write code, because the session starts with either too little context or too much. The hard part is not storing more instructions; it is deciding what the agent must load now, later, only on a trigger, or never by default.
The Challenge
Most agent setups treat context like a pile: project instructions, operating manuals, safety rules, old incidents, runbooks, and task notes all get shoved into the launch prompt because each one once mattered. That works for a while, then collapses under its own weight. The model pays for stale history on every session, active rules get buried beside background reference, and the human cannot easily see which sources were actually required for the work at hand. Semantic search helps with discovery, but discovery is not authorization; a fuzzy match is not a safety gate.
The Approach
Smart-Kickoff treats startup as a load contract. Instead of asking an agent to read everything before doing anything, it separates context by timing: what is required now, what is required before mutation, what loads only when a deterministic trigger fires, what can be found by semantic retrieval, what belongs in a deferred runbook, and what is optional reference. A resolver checks the initiating work memo, writes a manifest of the sources and load classes, and refuses large undeclared context that cannot provide a compact launch slice.
What Changes
Before: every session pays for full manuals whether the task needs them or not, and the agent can follow the reading list while still missing which facts are active. After: the agent starts with a compact orientation, state-changing operations trigger exact safety sources, and the manifest makes the launch package reviewable. The result is not less context; it is context loaded at the moment it earns its cost.
Go Deeper
- Load contracts — separate launch context from mutation gates, lookup, runbooks, and optional reference.
- Mutation gates — require exact sources before actions that change code, data, credentials, deployments, or shared memory.
- Manifest receipts — record what loaded and why without embedding private source bodies in the prompt.
We built this because “read everything first” felt safe but behaved like a tax. The better pattern is stricter and lighter: make every source state why it is present, then make the system fail closed when that claim is malformed.