Let’s be honest - portfolio monitoring is not what most dealmakers wake up to. Most medium-sized funds are under-resourced for it. Between the excitement of originations and the tedium of stress-testing loans and portfolios across a range of macro and company-specific conditions, most in the industry would rather do the former.
The sector, in the Financial Stability Board's words, “remains untested to a prolonged economic downturn.” This is the setting in which open source AI agents matter - not mainly as a productivity tool for funds, but as a new instrument for the people whose job is to keep the system safe.
What we learned building agents for portfolio monitoring
We began building AI portfolio monitoring for several private credit funds, using accessible tools from the frontier labs. That made it easy to tailor the monitor assessment methods in the way the firm actually collaborates. AI let us run the kind of deep, repeated stress testing that resource constraints had always made impractical, and in our backtests it caught several red flags that turned out to be meaningful value indicators months ahead.
Once set up, the agent could go deeper into each loan than a stretched team ever could. It can raise red flags for a borrower seemingly shuffling inventory to meet its covenants. Partners could ask it to forecast across a range of scenarios at the deal and portfolio level - which loans are most exposed if gas prices rise 10 percent, say - and get in minutes what used to take days.
The most important design choice was to keep analysts at the centre. Their involvement in setting up the monitor was what took the AI’s output from something partners judged 50 to 60 percent “accurate” - meaning how well the generated deal memo reflected what those partners would have flagged - to well above 90 percent. Building the right checkpoints for analysts at setup was decisive for both quality and adoption. Deals are heterogeneous, and much of the judgment that makes them so - the adjustments to free cash flow or EBITDA for a particular company or deal type or deep contextual awareness of the borrower - sits in analysts’ heads. Once captured, much of the recurring work could move onto the AI workflow, freeing analysts for higher-value engagement with borrowers.
A related and less glamorous piece of the work is data. Many firms have limited ability to build data pipelines, and they work largely with unstructured and semi-structured material - legal documents, industry databases, spreadsheets. This is where the current generation of AI genuinely excels. Traceability matters just as much: trust is built by letting everyone see where each number was derived from.
Why open source, and what stays in-house
For a fund, the case for open source begins with speed and ownership. The hard part of portfolio monitoring is never the model that reasons over a loan - the frontier labs supply that - but the scaffolding around it: the definitions, the analyst checkpoints, the sequence of steps that turns a pile of reporting packages into a deal memo a partner will trust. Building that from scratch is exactly what a stretched mid-market team has no time for. A shared, open set of monitoring and covenant agents hands a firm that scaffolding on day one - a working foundation it can deploy in a day rather than quarters - and, because the agents are open, it can be configured to your company's processes instead of you inheriting someone else's.
The reflexive objection to any AI in this seat is that it is a black box: a partner is asked to sign off on a number without being able to see how it was reached. This is precisely the objection open source dissolves. Proprietary tooling asks you to trust the vendor; an open agent lets you read it. Every definition it encodes - what counts as an EBITDA add-back, how a covenant is specified, what a stress scenario contains — is visible, inspectable and auditable, and every figure it produces traces back to the document it came from. Trust here is not asserted, it is shown instead. The same traceability that let our partners move from judging the output 50 to 60 percent accurate to well above 90 percent is a property of the workflow being open, not a feature bolted on top.
Being open also means the process stays yours to change. A vendor's black box hands you its template and its judgment; an open agent hands you the source. The adjustments that make a given deal your deal are yours to rewrite, not settings you are permitted to toggle. The common layer gives everyone a shared vocabulary of credit standards; what each firm builds on top of it is proprietary, because that is where your edge lives. Owning your core process is not a nice-to-have - for a credit fund it is the whole point.
It is worth being exact about what "open" refers to, because the instinctive fear is that open source means opening the book. It does not. What is shared is the workflow - the agent, the scaffolding, the common definitions - never the data or the alpha. In fact the open workflow's most valuable job is the opposite of exposure: it is the mechanism by which a firm captures the judgment that today lives only in its analysts' heads - the add-backs, the context on a borrower, the reasons a covenant was written the way it was - and stores it, structured and reusable, inside the firm's own systems. The agent runs against your book, on your infrastructure; nothing about the deal ever leaves. The shared infrastructure is not the edge. The firm's context is - and open source is precisely what lets you accumulate it without giving it away.
The time is ripe
Open source agents and agentic workflows for finance firms that live and breathe generating alpha might seem radical today. Software first embraced open source in the 1980s and the idea that giving code away could build stronger products and industry was equally radical then. Finance is at a similar moment with this generation of AI technology. New platforms are being adopted across the sector at remarkable speed, reaching non-technical partners and analysts alike, which means the base on which shared agents could spread already exists. Used well, they can put a more robust risk toolkit in every firm’s hands, upskill a sector that runs on tradition, and let standards disseminate rapidly and effectively.
Open source AI agents will greatly help resource-strapped private credit firms navigate our uncertain times. Open sourcing helps disseminate these tools more quickly without long setup time, extensive demos and unverifiable sales pitches. Try it here and let us know how we can make it work better for you.