What Actually Are Agents?

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

…making them investors in all three frontier AI companies. Another sign of how different this cycle is.

David Solomon sat down with TIME this week and explained what increased productivity for Goldman actually means. Hint: it's not just doing more with less.

And this week, I want to talk about what agents actually are. Because I think most people have it wrong.

In This Week’s Issue:

From The Trenches

What Agents Actually Are (And Why It Matters)

There's a fundamental misunderstanding about what AI agents are going to be capable of. Most people think of them as fancy chatbots. Research assistants that can look things up faster. That's not where this is heading.

Agents will be able to execute entire workflows end to end. Not "summarise this document" or "find me news about this company." Actually do the work. Receive a CIM, extract the financials into a structured format, build a model using your firm's template, benchmark against comps, flag the risks, draft the IC memo. The whole thing.

This is why agents are going to be far more valuable than people realise. The good ones won't be generic tools you pull off the shelf. They'll be trained specifically on relevant context. Your firm's processes. Your templates. Your investment criteria. Your way of structuring a memo.

Just like you might have an associate who's particularly strong on healthcare deals and another who's better at industrials, you'll have agents with different specialisations. Different training. Different strengths. The difference comes down to the context they've been given, the instructions they've been given, and who's trained them.

The Analyst Parallel

This reminds me so much of my analyst days, actually. When I started at J.P. Morgan, nobody handed me a manual and said "go do deals." I learned by looking at previous examples and replicating them. You find a model someone built for a similar transaction. You study how they structured it. You try to do the same thing. You get feedback. You iterate. Over time, you develop judgment about what good looks like.

That's basically what training an agent is like. It does work, you provide feedback, it improves. The parallel is almost exact. The agent learns from examples. It replicates patterns. It gets corrected when it's wrong. It develops something that looks a lot like judgment over time.

But two important differences.

One, they don't get tired in the traditional sense. The investment is evergreen. But two, the agent has to have the right "tools" to be able to do its thing. I can reason about a lot of things, but if I can't actually do something, the reasoning is just noise. The gap between "understanding what needs to happen" and "being able to make it happen" is where most agent projects struggle. You need both the intelligence and the capability to act.

What do I mean by tools? Tools are the capabilities that let an agent actually do things in the real world. Read a file. Write to a database. Send an email. Update a CRM. Pull data from an API. Without tools, an agent is just a very articulate thinker with no hands.

The more tools an agent has access to, the more it can do. But each tool has to be built, tested, and maintained. And the tools have to work reliably with messy, real-world data. That's where the complexity explodes.

This is why we've spent so much time at DealSage on things that sound basic. Take something we do as humans that feels straightforward but is really hard for agents: reading and writing to Excel. Excel is a nightmare for agents. It's designed for humans to look at, not for machines to parse. A human sees a blank row and knows it's a section break. A human sees "Revenue" in column A and knows the numbers to the right are revenue figures. An agent has to be explicitly taught all of that. Every firm's templates are different. Every banker has their own quirks. Making an agent Excel-native is genuinely hard. We've spent months on it.

Why Building Agents Is Hard

I saw a post on LinkedIn this week from a product manager at Beehiiv. He'd spent 90 days building an AI agent into their platform. Three months. For a product that is, at its core, a fancy text editor. And they still haven't deployed it.

His takeaway: "Building an agentic product isn't just another feature launch. It's a total mindset shift and has a pretty brutal learning curve."

People are massively underestimating the work that goes into creating good agents.

The Compounding Effect

An analyst takes two years to become genuinely useful. Then they leave for business school or a different job. The knowledge walks out the door. You start over with the next class.

An agent, once trained, doesn't leave. It doesn't forget. You can clone it. The institutional knowledge compounds instead of walking out the door every two years. That's a structural shift in how firms build capability over time.

What This Means

Right now, everyone's talking about "building their own agents." My prediction: that won't scale. Just like the job market, you'll hire agents with specific skill sets that can be plugged into your workflows and data.

Think about why you recruit analysts from the best schools. It's not just intelligence. It's that they've been taught well. They've learned frameworks. They've been trained on what good looks like. The same will be true for agents. You'll want agents that come from teams who've trained them properly.

I don't see a world where everyone is building, training, and maintaining all their agents internally. The training is too hard, too time-consuming, too specialised. You'll "hire" agents the same way you hire employees today. Except they'll actually do the work consistently and won't leave after a couple of years :)

At DealSage, we're building an agent that can actually do deals. Not just talk about them. That means giving it the tools to act, not just the intelligence to reason. It means training it on what good looks like in M&A specifically. It means iterating constantly as the technology evolves.

The future isn't AI that helps you do your job. It's AI that does the job, with you managing the output. That's a different world. And it's arriving faster than most people expect.

News Digest

Sequoia Breaks VC Taboo, Backs All Three AI Giants

Sequoia Capital is joining Anthropic's $25 billion funding round, according to the Financial Times. This makes Sequoia the first major VC to invest in all three frontier AI companies: OpenAI, xAI, and now Anthropic.

The move breaks one of venture capital's oldest unwritten rules: don't back competing companies in the same sector. Sequoia famously walked away from a $21 million investment in Finix in 2020 because it competed with Stripe, a Sequoia portfolio company. Now they're betting on three direct competitors at the same time.

The round values Anthropic at $350 billion, more than double its $170 billion valuation from just four months ago. GIC and Coatue are each contributing $1.5 billion. Microsoft and Nvidia have committed up to $15 billion combined. Total raise could exceed $25 billion.

Why it matters: When one firm backs all three frontier AI companies, they're either betting the market is big enough for multiple winners or they're just trying to hedge and don't know how this is going to play out.

My take: One source told the FT this doesn't feel like venture investing anymore. "It's a round where the company is so big that it's gone from a VC investment to a stock investment." That's the right frame. Sequoia isn't picking winners. They're buying exposure to the entire category. The AI market is so large that backing all three might be the rational move.

Read more

Goldman’s Solomon on AI: "It Frees Capacity to Invest"

David Solomon sat down with TIME ahead of Davos to discuss the economic outlook for 2026. The conversation turned to AI, and Solomon offered a frame that goes beyond the usual "efficiency" talking points.

"It's not just that that creates more margin accretion," Solomon said. "It actually frees up capacity to invest in things that you want to invest in to grow the business where you've been more constrained."

This is Jevons paradox in action. When you make knowledge work more efficient, you don't just do the same work with fewer people. You tackle bigger problems. You serve more clients. You enter new markets. Goldman isn't maintaining a fixed amount of work with fewer analysts. They're expanding what's possible.

Why it matters: The narrative is shifting from "AI replaces jobs" to "AI unlocks capacity." The firms with the biggest backlogs of ideas they never had time for will benefit most.

My take: Solomon's point about freeing capacity is the right one. The question isn't whether AI will eliminate roles. It's whether your firm has enough ambition to deploy the capacity it creates. The firms thinking "how do we do the same thing with fewer people" are playing defense. The firms thinking "what couldn't we do before that we can do now" are playing offense.

Read more

Hyperscalers to Borrow $140B Annually as AI Becomes a Credit Story

U.S. corporate bond issuance is expected to hit $2.46 trillion in 2026, up 11.8% from last year. The biggest driver? AI hyperscalers funding their infrastructure buildout through the debt markets.

The Big Five hyperscalers (Amazon, Alphabet, Meta, Microsoft, Oracle) issued $121 billion in corporate bonds in 2025. That's versus an average of $28 billion per year between 2020 and 2024. A 4x increase. BofA expects them to borrow roughly $140 billion annually over the next three years, potentially exceeding $300 billion.

Meta's $30 billion deal in October was the largest-ever individual non-M&A high-grade bond sale. Four of the five biggest U.S. high-grade bond deals in 2025 came from hyperscalers.

The details:

Why it matters: AI is no longer just a venture capital or equity story. It's becoming a credit market story. The hyperscalers are borrowing at a pace that could make them as big as the major banks in the investment-grade index.

My take: The Oracle lawsuit is interesting. Bondholders claim the company failed to disclose it needed to sell significant additional debt for AI infrastructure. That's going to become a recurring theme. Everyone's racing to build out AI capacity. Not everyone's being transparent about what it costs. When hyperscaler CDS spreads are tripling, someone's worried about the bill coming due.

Read more

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

Oracle bondholders sue over undisclosed AI debt needs (Jan 15) - Claim company failed to disclose it needed significant additional debt for AI infrastructure. Oracle's five-year CDS has tripled since September. When the bond lawyers start circling, pay attention.

ElevenLabs eyes $11B valuation for voice AI (Jan 18) - The voice AI startup that powers half the podcasts you listen to is raising at unicorn-plus territory. Your earbuds are about to get a lot more synthetic.

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