How Agents Actually Work

I round up the most relevant AI-in-finance news - the deals being done, who’s rolling out what, and what’s actually working on the front lines

…and the Dow dropped 800 points. Citadel published a rebuttal. And then, three days later, Jack Dorsey announced Block was cutting 40% of its workforce, explicitly citing AI. "Intelligence tools have changed what it means to build and run a company." The stock rose 14%.

Jensen Huang weighed in too. Nvidia's CEO pushed back on the SaaSpocalypse narrative, arguing that agents won't replace software. They'll sit on top of it. The infrastructure for this new world is being built in real time.

But first, my take on what all of this actually means in practice. Because there's a new vocabulary emerging around AI, terms like API, MCP, and CLI, and I think understanding it is about to matter a great deal for anyone in deal-making.

In This Week’s Issue:

From The Trenches

How Agents Actually Work

ServiceNow admitted on its last earnings call that "agentic workflows" were complicating the long-term visibility of seat-based growth. Salesforce is down 26% year to date. The iShares Software ETF is off 20%. The "seat-count crisis" isn't a theoretical risk anymore. It's showing up in quarterly results and contract renewals, as enterprises realise that AI agents doing the work means fewer humans needing software licences.

That's the macro picture. But I think most of the debate is skipping over something more fundamental: how do agents actually do any of this?

Not philosophically. Mechanically. How does an agent pull data from one system, process it, and push results somewhere else? How does it connect to your email, your data providers, your deal platform?

There's a new vocabulary emerging. API. MCP. CLI. You're hearing these terms in every conversation about enterprise AI right now. If they don't mean anything to you yet, they will soon.

API stands for Application Programming Interface. In plain English: it's the way one piece of software talks to another.

When you log into PitchBook and pull up a company profile, you're using their interface. You click buttons, scroll through data, read charts. That interface was designed for you. A human. An API is the equivalent, but designed for software. Instead of clicking a button to get a company's revenue, an agent sends a structured request and gets back structured data. No screen. No clicks. Just a direct connection.

Every tool you use has an API, or should. Your email. Your CRM. Your data providers. Your deal platform. When people say a product is "API-first," they mean it was built so that other software, including agents, can interact with it natively.

This matters because an agent is only as useful as what it can connect to. An agent without APIs is a very articulate thinker with no hands.

And this is why seat-based software pricing is in trouble. If your product was built for humans to click through, and it doesn't also have an API that agents can call, it's invisible to the new workflow. The agent can't use it. It routes around it. The products that survive will be the ones that work for both humans and agents.

Here's the problem. Every API speaks a slightly different language. Gmail's API works one way. FactSet's works another. Salesforce's works another. If you want an agent to connect to ten different tools, you historically needed ten different custom integrations. Each one fragile.

MCP, the Model Context Protocol, is an open standard that solves this. It creates a universal way for AI models to connect to external tools and data sources. One protocol. One standard. Any tool that supports MCP can be plugged into any agent that supports MCP.

Think of it like USB. Before USB, every device had its own proprietary connector. Printers, keyboards, cameras, all different cables. USB standardised the connection. MCP is doing the same thing for AI.

This is why Anthropic's announcement this week matters. They released MCP connectors to FactSet, MSCI, and S&P Global. Agents can now pull institutional-grade financial data through a standardised connection. The same protocol that connects to your email connects to your market data. And because MCP is an open standard, any agent, any model, any provider can use it.

I ran a demo last week. One window. From Claude, I pulled recent emails from Gmail related to a deal we were reviewing. Sent the context to DealSage, which had the CIM and supporting documents already structured. DealSage's agent extracted the audited financials, ran benchmarking against comparable deals in our database, and pushed the structured output back into Claude. I combined everything into a draft IC memo. One session. No switching tabs.

Every step worked because each tool exposed an API, and MCP provided the standard connection layer. Gmail, DealSage, the financial data, all plugged into the same agent through the same protocol.

I don't want to pretend this is frictionless. I've spoken with many frustrated VPs at PE firms recently who want to connect their AI tools to Dropbox, their data providers, their internal systems, but can't. Compliance hasn't approved it. IT hasn't scoped it. Sound familiar?

Security and access controls are legitimate concerns. But I think the bigger bottleneck over the next 12 months is governance. Specifically, lineage and sourcing.

When an agent pulls a gross margin number into your memo, where did that number come from? Which document? Which version? Was it the audited financials or the management case? If you can't trace the answer back to its source, you can't trust it. And if you can't trust it, you can't put it in front of an investment committee.

The answer isn't controlling the model. It's controlling the data layer. Making sure every piece of information that flows through an agent workflow has a clear source, a clear timestamp, and a clear chain of custody. That's what we've been building at DealSage, and it's one of the reasons we built with MCP from the start.

The firms that figure out connectivity first will compound. Every new tool you connect, every new data source you plug in, makes the agent more capable. The firm that connects five systems this year will be running circles around the firm still copy-pasting between tabs in 2027.

I think every piece of software needs to work for both humans and agents. The products built only for human eyes, the ones charging per seat for a dashboard with no API, are the ones in trouble. The products that expose structured data and agent-friendly connections become the infrastructure of the new stack.

I'm going to do a full piece on what that stack actually looks like in a coming issue. What the investment firm of the future is built on, how the pieces fit together, and what it means for how firms are structured, staffed, and valued.

For now: API, MCP, CLI. Learn the vocabulary. Understand the connections. The firms that wire up first won't just be more efficient. They'll be playing a different game entirely.

News Digest

Block Cuts 40% of Its Workforce in AI-Driven Remake

Jack Dorsey announced on February 27 that Block would eliminate roughly 4,000 positions, approximately 40% of its workforce. The letter to shareholders was unusually direct. "The intelligence tools we're creating and using, paired with smaller and flatter teams, are enabling a new way of working."

Block's stock rose nearly 14% the following day. The market didn't see this as distress. It saw it as discipline.

The details:

Why it matters: This is the first time a major public company has cut at this scale and pointed directly at AI as the reason. When a $45 billion company says it can operate with 60% of its headcount because of AI, every PE portfolio company's headcount assumptions need revisiting.

My take: There’s probably more to this than initially meets the eye. Block ballooned from about 3,800 employees in 2019 to over 10,000 during COVID. Dorsey himself admitted on X that he overhired and "incorrectly built 2 separate company structures." Goldman noted the reduction takes headcount back to 2020 levels. An Oxford Economics report released in January found that many layoffs CEOs called AI-related were actually the result of past overhiring. So is this genuinely AI-driven transformation, or is AI a convenient narrative to dress up a long-overdue correction? Probably both. Dorsey is using real AI capabilities to justify a restructuring that was needed regardless. But the market reaction is what matters. Investors rewarded the decision. They didn't care about the reason. They cared about the margin improvement. And that's the playbook every PE operating partner is now studying: whether the AI thesis is 100% genuine or partly opportunistic, the market is rewarding lean.

Read more

The Citrini Memo and the Week the AI Scare Turned Real

On February 22, Citrini Research, the most-read finance newsletter on Substack, published a 7,000-word essay titled "The 2028 Global Intelligence Crisis." Framed as a memo from the future, it described a world where AI displaces white-collar workers en masse, the S&P 500 falls 38%, unemployment hits 10.2%, and the economy enters a deflationary spiral. The memo introduced a concept it called "ghost GDP": economic output that benefits computing power owners but never reaches the consumer economy.

The Dow dropped over 800 points on Monday. Software stocks were hit hardest.

By Wednesday, Citadel Securities had published a rebuttal. Software engineer demand was up 11% year over year. Daily use of generative AI at work was "remaining unexpectedly stable." The core flaw in Citrini's logic, Citadel argued, was assuming "recursive technology equals recursive economic adoption." Physical constraints, cost thresholds, and human adaptability would brake the acceleration.

Then Block made its announcement. And the debate shifted from theoretical to very real.

The details:

Why it matters: A Substack post moved the Dow. That tells you something about the level of anxiety beneath the surface. Whether Citrini's timeline is right or wrong, the fact that markets reacted this violently to a thought experiment means the uncertainty is already priced in as risk.

My take: Citadel's rebuttal was well-argued. The economics in Citrini's piece are speculative and probably too aggressive on timeline. But three days after Citadel said there was "little evidence of any imminent displacement risk," Block cut 40% of its workforce and the stock went up. The data says one thing. The decisions being made by actual companies say another. The truth is moving toward Citrini's direction faster than the macro data captures.

Read more

Anthropic Launches Financial Services Plugins for Claude

Anthropic announced on February 24 a suite of financial services plugins for Claude Cowork, with dedicated tools for financial analysis, equity research, investment banking, private equity, and wealth management. The PE plugin supports deal sourcing and diligence by reviewing document sets, extracting financial data, modelling scenarios, and scoring opportunities against investment criteria. They also released connectors to FactSet, MSCI, S&P Global, and LSEG.

Anthropic continues to cement itself as the frontier model of choice for knowledge work. This is the same model powering Goldman's embedded agent programme. The same one that triggered the SaaSpocalypse sell-off. And now it's shipping finance-specific tooling out of the box.

The details:

Why it matters: Anyone who read our orchestration issue a few weeks back will recognise the pattern here. Central interface, connected data sources, agents that execute tasks. Anthropic is laying the template for exactly that workflow.

My take: For solo practitioners or small teams without dedicated tech resources, this is a genuinely compelling option. Pull some market data, run a comp, get a quick read on a target. Available today, no engineering team required.

We've long expected that point analyses would commoditize. This is that happening. And it's a good thing. The ability to do all of this has technically been there for a while. Anthropic has just packaged it in a way that's made it more accessible as an off the shelf option. The value, at least where we sit at DealSage, is in what sits alongside the analysis: the repository, the deal history, the lineage, the institutional brain of your organisation that compounds over time. The firms that build the memory around it will pull ahead.

Read more

Other Interesting Things I’ve Read of Seen This Week:

KKR private credit fund reports jump in troubled loans (Feb 27) - FS KKR Capital Corporation, a $13B vehicle, lowered its return guidance after marking down several positions. The private credit stress story from last issue isn't getting better.

Google strikes multibillion-dollar AI chip deal with Meta (Feb 27) - Meta signed a multi-year deal to rent Google's TPUs for AI model training. When the two biggest social media companies start cutting chip deals with each other, the AI infrastructure arms race has entered a new phase.

Bank stocks suffer another plunge on credit and AI fears (Feb 27) - Banks and brokerages fell sharply as investors repriced AI disruption risk alongside private credit concerns. The "AI scare trade" from the Citrini memo is still rippling through.

Rathbones CEO outlines AI-driven adviser productivity strategy (Feb 27) - New CEO plans to stem client outflows by using AI to boost adviser productivity and cut servicing costs. Not replacing advisers. Making each one more effective. The Jevons paradox playbook applied to wealth management.

Jensen Huang says the SaaSpocalypse narrative is wrong (Feb 26) - Nvidia's CEO argues agents won't replace software. They'll sit on top of it. The value accrues to the platforms agents use, not the agents themselves. Convenient framing from the company selling the picks and shovels.

Acquisition Intelligence is a weekly newsletter on AI in M&A for finance professionals, private equity investors, investment bankers, corp dev teams, and deal-makers.

For questions, feedback, or to share what you're seeing in the market, reply to this email.

P.S. I'm Harry, co-founder of DealSage. We're building an AI-native deal intelligence platform to help professionals turn their institutional knowledge into better decisions. If you're curious what we're up to, check out dealsage.io or just reply here