naïvelabs
AI is only as capable as theinfrastructure it’s built on.
We research the agentic layer AI runs on: how far frontier agents can scale without a machine per seat, and whether shared context can compound across a company.
[Papers]
Papers clear the research gate before anything is published. Soft launch: one titled research direction (abstract only) — no PDF until the gate admits it.
- 7/25/2026Toward denser sandboxes: isolate runtime research notesResearch directionResearch direction on denser sandboxes (V8 isolates). A task-scoped disposable box is one product shape under study — not a shipping SLA. Soft launch: no PDF until the research publish gate clears.PDF after research gate
[Harnesses]
06
Instruments we care about for agent infrastructure. Soft launch: working titles and intents only — not a published results board. Numbers live in papers once they clear the research gate.
- 01Long-Run RecoveryORCH
Whether long serverless runs survive interruption
- 02Graph EscalateORCH
When multi-agent work should stop and ask a human
- 03Recall Under DriftBRAIN
Brain retrieval as company context changes
- 04Consolidation DecayBRAIN
Whether shared context sharpens or just grows
- 05Cost Per Box-SecondSBX
What a disposable box costs for the time it runs
- 06Route RegretINF
Cost of the model pick versus better available options