Blog indexTagged
2026
A developer gives a project eight focused hours a day; the other sixteen the codebase sits idle. A good harness moves work into those hours: interval-scheduled agents on a remote machine running refactors, idea-to-task pipelines, and bug hunts on separate branches — timed so they never collide. Here's the philosophy and the math.
ReadSpectral's orchestrator never does the work itself — it delegates to researcher, executor, reviewer, and fork subagents in waves. Each wave is a parallel call plus a barrier where all reports are verified and distilled into a few lines that feed the next wave.
ReadIn-process subagents are cheap but invisible. Spectral's session_fanout tool turns a batch of independent research tasks into real, persistent sessions that run in parallel — with caps, semaphores, and recursion guards keeping it safe.
ReadSpectral's cross-session project memory used to match exact substrings only — multi-word queries and typos returned nothing. Here's how n-gram indexing, ranked scoring, and near-duplicate detection make observational memory actually useful during development.
ReadLarge tasks don't have to be slow or expensive. Here's how Spectral's prefix caching, byte-stable prompt, and session affinity let it reuse ~97% of prompt tokens from cache — making big multi-turn agentic sessions fast and cheap.
Read