AgentKV — KV Cache Control Plane for Agents

Let developers manually manage inference KV cache when forking agents and running long tasks

Problem

Developers report that black-box inference providers leave them without manual control over KV cache. For agent swarms, forking multiple agents from the same context is especially painful because they are stuck with opaque caching policies. This lack of control hurts efficiency and cost for long-running and branching agent workflows.

Solution

A lightweight inference proxy layer that supports manual KV cache forking, snapshotting, and reuse. Developers can fork multiple agents from a chosen context branch point or retain key cache segments for long tasks to reduce repeated computation and token spend. It integrates with mainstream inference APIs.

Users

AI engineers and indie developers building agent swarms, long-running agents, and fork-and-merge workflows

MVP

A proxy layer intercepting inference requests; cache fork markers and reuse; simple SDK and CLI to fork agents from a given context; OpenAI-compatible API; log cache hit and savings estimates

Revenue

Per-request relay pricing (API margin) or subscription for advanced enterprise features; open-source core proxy to drive adoption

Why now

Agent workflows are shifting from single-turn Q&A to long-running, multi-branch collaboration, but inference providers still keep KV cache control locked down; indie developers can ship a community-driven proxy layer faster than incumbents