Why I built AICEP: eight months, five agent harnesses, one handoff problem

How moving my own work between agent harnesses led to the AI Context Exchange Protocol.

Where it started

AICEP began with my own agent work, which kept outgrowing the tool it lived in. I name these tools only to describe the path I took, not to rank or endorse any of them.

In February 2026 I started with OpenClaw. In March I moved my day-to-day agent work to Hermes, and it stayed there through July. Through those same months I also ran Google AI and Claude subscriptions in parallel, so the same projects were being worked on in several places at once. In August I moved to Grok Bot, where my agents run today.

What each move showed me

Every migration, and every day of running harnesses side by side, meant moving context: project briefs, decisions, files, task status, and the standing instructions that told each agent what its job was. The parallel months made it sharper: when three harnesses touch the same work, each holds a different slice of the truth. Some context was lost along the way, some had to be re-entered by hand, and some was at risk of ending up where it should not. Secrets, private data and agents' roles had no portable rules. Whatever rules existed lived inside one harness and did not travel with the work.

A problem bigger than one migration

What I hit in my own moves is now common. Agents hand work across many harnesses and roles: one writes code, another files notes, another coordinates the rest. Those handoffs are where secrets, private data, user uploads and agents' job boundaries leak.

Agents hand off work across tools and roles. AICEP's rules keep that handoff clean: secrets, private data, and agents' jobs stay where they belong.

The founding aim

I wanted something simple to state. A new agent can be pointed at the protocol on setup to load its privacy and sharing rules, so context can move across clients and tools without leakage or inappropriate sharing. That aim became AICEP, the AI Context Exchange Protocol.

The through-line is interoperability without homogenization: lossless handoffs between tools, teams and generations, with human judgment kept in charge. AICEP is harness-neutral. It is meant to pass context in and out of the agent harnesses people already run, not to replace them.

My response, in four lines: Classify before you send. Seat an agent before it talks. Keep credentials off the bus. Exchange second.

A longer lineage

AICEP has a predecessor. I co-authored PhysBits 1.0, a physical-security interoperability standard, for the Open Security Exchange in 2003, and it was presented at ASIS, ISC West and RSA. AICEP takes many inspirations from this work, focusing not just on identity credential issuance between disparate systems, but full agent and context migrations between AI harnesses.

Where it stands

On 15 October 2026 I published AICEP 0.1, through AIOptiTech, as a Public Draft (RFC), open for comment. It is not a chatbot standard and not a certification. This first version is written primarily for individual users. The next revision takes AICEP from individual use to corporate use, incorporating more elements of corporate cybersecurity standards into the protocol.

Help shape it

I'd love to hear from anyone interested in helping develop this protocol or contributing to it. Read the Public Draft (RFC) at https://aicep.ai/protocol/, find the companion runbook at https://aicep.ai/runbook/, and send comments, implementation notes and questions to hello@aicep.ai.