
Nobody Sells The Whole Platform Any More
The requirement arrived as a compliance note: everything self-hostable, data resident in a room we control. So one platform became a list of parts. I read 14 candidates down to their licences, and the money had already moved somewhere I was not looking.
The Contract Asked For A Room
The requirement arrived as a compliance note, not a product decision. Defense work, ITAR and EAR, and one sentence that killed every hosted tool on the market: everything self-hostable, in San Diego or in the EU, with the data sitting in a room we control. The platform we had been using, Manus, worked fine. That was the whole problem. It was someone else's building. Six capabilities had to be replaced: agents, hosting, auth, analytics, a data API, and a proxy to the models. I opened a spreadsheet and put 14 candidates in column A, then started column B, which was licences. Choosing a product was never the job. Rebuilding one out of parts was the job. Fourteen candidates is not a short list. It is a sign that the thing I was shopping for stopped existing as one item.
Six Line Items, Not A Product
Manus looked like one product. It was six. Dify carries the agent builder, the hosting story, user management, logs and a REST data API, with 141,899 stars and 22,304 forks behind it. Docker Compose, Postgres, Redis, a vector store, and nothing phoning home. n8n brings 188,693 stars, 57,819 forks and 400+ integrations into the workflow layer, and it deploys as a single container. AutoGPT, at 184,429 stars, still has the rawest autonomy of the set. None of them is Manus. I kept hunting for the one row that would close the spreadsheet. There is no such row. The category stopped being a product a while ago and nobody sent the memo. The star counts are a popularity measure. They are not a compliance measure.
A Licence Can End The Adoption
The first three candidates died in column B. AutoGPT runs Polyform Shield, which a lawyer reads as a commercial restriction. n8n runs the Sustainable Use License, fair-code rather than open. Open WebUI limits branding removal at 50 users. Dify ships modified Apache 2.0, and the modification is the only part a defense programme cares about. The safely permissive ones — CrewAI, LangChain, LiteLLM, smolagents, Agno, AGiXT, all MIT or Apache 2.0 — mostly are not platforms at all. The probability that a licence drafted for a cloud business quietly disqualifies a defense deployment is not zero. Most buyers price it at zero. I read licences before feature lists now. It costs an afternoon. It is cheaper than a rewrite.
A Framework Is Not A Platform
CrewAI is 51,725 stars, MIT, and the cleanest multi-agent role-play framework I have used. It also has no hosting, no auth, no analytics and no data API. That is not a defect. It is a different product wearing the same word. LangGraph, at 32,416 stars, brings stateful multi-agent orchestration, and the honest estimate for everything around it — interface, login, logs, API — was three to six months of engineering. I have built companies in twelve countries, and the gap between the demo and the deployment is always the same width. The framework works beautifully in the room. Then you discover that you are the platform. Budget the other four pieces before you fall for the one you like. Three to six months is not a warning. It is a quote.
An agent platform is now a list of installable components. No vendor owns the list.
The Piece Nobody Budgets For
Every architecture in that document ended with the same component, and it was not an agent platform. LiteLLM is a proxy: 47,565 stars, MIT, 100+ model APIs behind one OpenAI-shaped endpoint, with keys, cost tracking and routing. It is the only place in the stack where you can decide that no prompt leaves the building. Block the external providers. Route to Ollama or vLLM inside the perimeter. Swap the agent platform and the proxy stays. That asymmetry deserves attention. The component with the smallest surface area turned out to be the one the compliance review actually cared about. On my first pass I had it at the bottom of the list. It should have been at the top. A proxy is boring on purpose, which is exactly the property you want in the component everything else must pass through.
Eighty Per Cent, And The Margin Moves
The recommendation came out as two containers, not one platform. Dify for agents, auth, logs and the data API, LiteLLM for the proxy, roughly 80% of Manus covered and the remaining 20% as custom work. The fallback if Dify's terms fail review is n8n plus Open WebUI plus the same proxy. That is the finding. An agent platform is now a list of installable components, and no vendor owns the list. Which moves the money. Margin used to sit inside the subscription. It now sits in the integration, the licence review, and the four pieces you have to hold together yourself. The rent went down. The work went up. One vendor used to carry all of it. Now four do, and the seams between them are mine.
The probability that a licence drafted for a cloud business quietly disqualifies a defense deployment is not zero. Most buyers price it at zero.
The map is dead. Nobody told you.
Bali State of Mind is the survival guide for the collapse of everything you were taught to believe.
Beyond this book
Building the same thing somewhere else.
Julien Uhlig is available for advisory work, board seats and media appearances. Write to media@exventure.co.
The academy that trains the operators, across every company in the group, is EX Epic Academy - 25,000 applications, 25 seats per cohort, 210 alumni across 19 countries. academy.epicsolutiongroup.com
18-20 November. Online, Las Palmas, Bali.
Three days on what happens to work, capital and institutions when the map stops matching the ground. Seats are limited by cohort.
ex-aisummit.com →