Opening fund operations to agents

Operating a venture fund is a coordination effort between a fund's managers, their limited partners (investors), their portfolio companies (investments), and a mix of internal resources or third-party entities that manage banking, fund administration, audit, tax, and compliance. Today these conversations run through email, portal logins, shared spreadsheets, and manual communication between these parties.

As our workflows move increasingly into chat interfaces, we are seeing efforts to make data available via connectors, plugins, MCP apps, et. al. to help us access data without logging into websites and do our work using that data inside our LLM chat interfaces. That first step is being quickly followed by increased use of automations and agents to do that work for us. People are using business and personal assistants like Instinct, Muse, Dots, Claude Duet, Grok, ChatGPT, Claude, Town, and more through interactive conversations and scheduled actions to help them do their work, and companies are increasingly building the infrastructure to enable them to work with agents efficiently.

A couple weeks ago I updated Portfolio by Hemrock to work as a headless fund administration app, an effort to help people do the same things via an LLM interface or API as they would by using the website. This week I updated the app to extend that same ability to agents, working towards building an agentic fund operations platform. Anything a person can do in the app, an agent can do, limited to what that person is allowed to see and do, powered by your own MCP server. ¹

What agentic fund operations means

Right now, fund managers and third-parties can use Portfolio to do a range of activities:

• The fund administrator can pulls trial balances and financial statements for any vehicle and period, and draft or post journal entries to update the financial statements.
• The auditor can read the ledger and check the audit trail without asking for exports.
• The tax preparer can get K-1 inputs, tax adjustments and a tax package for the year.
• Portfolio companies can submit their updates, decks and KPIs, and the app will update the firm's records and analyze the company's performance.
• LPs can get capital call notices, access their current performance metrcs, and pull capital account statements.

The goal is to enable agents to do all of this for people. I use "agent" loosely, to refer to anything that acts for a person, including:

• A scheduled task in someone's LLM. "Every Monday at 8, check my funds for anything due this week."
• An LP's personal or business assistant that can get performance information and do analysis for them.
• A scheduled API call from a provider's own system, running nightly or at quarter end, to get updates for audit, tax, financial performance, compliance, or more.
• An agent a provider builds into their product for every fund client they serve.

Examples of how to use it

Here's a few examples on how firms could set this up for people to use.

• The LP on a Sunday night. An LP in six funds asks their AI assistant, "Do I have any capital calls due this month, and did my last wire land?" The assistant checks each manager's Portfolio and answers: one call due on the 15th, and last quarter's wire was received. The LP says, "Put the wire on my calendar and save the notice to my tax folder." Nobody at either firm had to answer an email.
• The fund administrator at quarter end. The administrator sets a scheduled task to run on the first business day after quarter end. For each vehicle it serves, the task pulls the trial balance, compares it to the administrator's own ledger, and drafts the adjusting entries, optionally using any Skills or instructions that help define the process, checks, and details to make sure it runs as expected and deals with any issues. The drafts land in Portfolio as proposed entries, waiting for the GP to approve. The administrator gets one summary of what didn't tie.
• The tax preparer in February. A preparer who works with a dozen funds asks their agent to collect K-1 inputs and the tax package from each one. The agent comes back with a list: ten are ready, and two are still waiting on K-1s from funds they invest in. The preparer knows who to chase before anyone asks.
• The portfolio company CFO after close. A CFO has a monthly task that runs when the books close. It sends the board deck and KPI sheet to each investor's Portfolio. The figures land on the company's record at every fund that holds it, and no one at those funds sends an email, not to mention retypes cash or ARR into a spreadsheet to send.

Throughout all of this, the firm stays in control. Access is scoped separately for each entitiy and person. Every agent acts as a signed-in member. Changes are staged or drafted, not applied or approved unless that right is explicitly granted by the form. Everything that happens is tracked in an audit trail that can be checked end-to-end by the firm.

MCP is an open protocol, so a provider isn't tied to one model or one assistant to reach a fund. Write software that uses an API, add a connector to ChatGPT, add access to Instinct, the app will work with today's and tomorrow's set of agents as we continue to evolve the technology.

And of course, the app is open source, so if you would prefer it to work or look differently for you, you are free to change it to fit your desired workflows or situation.

If you're a fund manager, try connecting your own assistant to your fund and ask it who still owes on the last call. If you're an administrator, auditor or tax preparer who works with venture funds, we'd like to hear what your agents need to ask. And if you're an LP, ask your managers when your assistant can talk to theirs.

If you're interested in chatting more about the vision, how you can deploy this platform, or if you're inspired to build your own, contact me, I'm happy to chat. It's a great time to be building.