Sparkco AI Sheets

What is Sparkco AI Sheets?

Sparkco AI Sheets is an open-source research project exploring the post-AGI stack — the runtime containers agents live in, the glue code inside them, and the messaging between them. Built by the team behind SimpleFunctions, Sparkco investigates how live prediction-market probabilities can serve as a real-time world state for AI agents. The platform serves as a public log of this work, featuring a live feed of AI and prediction-market signals, along with curated tools and setups for agent builders.

Key features include CLI-first tooling that is pipe-composable, 0-token exposure, and ~100% reliability — a departure from MCP-based approaches. Sparkco addresses the fundamental limitation of current chat models: context windows that are finite token sequences, fully recomputed on every turn. Existing workarounds like summarization memory, vector retrieval, and KV caching move the cost without solving it — long context drifts, compaction discards information, and retrieval crowds the same window it pulls from. Sparkco tests whether conversational state can be encoded directly into model weights as weight deltas, leaving the context free for the current turn only.

Use cases include building autonomous agents that maintain coherent multi-turn conversations across technical, strategic, philosophical, and personal domains without context degradation. The project also explores durable runtimes, microVMs, and sandboxes where agents live, with SimpleFunctions providing autonomous daemons, a scheduler, and risk gates for prediction-market agents. Technical details involve comparing weight-based memory approaches across disjoint communities — efficient inference, recsys, personalization NLP, on-device, and model editing — on a unified benchmark using real user histories spanning hundreds of turns. This is a novel evaluation not covered by existing synthetic benchmarks like RULER, needle-in-haystack, or LaMP. Sparkco is ideal for researchers and builders pushing the boundaries of agent memory and context management.

Use Cases

Key Features

CLI first tooling pipe composable commands token exposure ~100% reliability live prediction market signals weight delta encoding context free architecture agent runtime containers public research log curated agent tools

Who is it for?

AI agent builders, research engineers, post, AGI developers, prediction, market analysts, open, source contributors, infrastructure engineers, machine learning researchers

Similar Tools