Insights

Laying the Foundation for Autonomous Private Capital. Step Three: Capturing Invisible Liquidity.

This is Part 3 of SecondLane’s series on the autonomous agents rebuilding the infrastructure of private capital. In Part 1, we learned to meta-prompt. In Part 2, we built the autonomous infrastructure. Now, we turn on the machine to capture the invisible liquidity.

Most companies treat AI like a highly efficient intern with amnesia. You ask a question, it answers, and then it forgets everything.

This approach is fine for drafting emails, but useless for discovering prices in a fragmented market. To capture alpha in private markets, we need AI that remembers. We need an architecture designed for “spatial awareness.”

The alpha in secondary markets is hidden in thousands of fragmented Telegram chats, Discord threads, X posts, and emails. A human trader can’t process this volume. They are trading with tunnel vision, often with a lag

The solution is to move beyond chatbots and build systems that ingest the market vibes and structure them into a high-confidence shadow order book.

Here is the blueprint for turning intent mentions into hard pricing signals.

Building the 4-Stage Agentic Loop

The future of the AI-native firm is about two things: using the right data and building an effective loop. In the heart of the loop lies the continuous 4-stage cycle:

The Memory Layer
This is the foundation. If your AI cannot access your history, it cannot have context. A solid system builds on a RAG-like ingestion of every internal data point: Slack conversations, email threads, pitch decks, and historical transactions. This gives the agent a “second brain.” It knows who the clients are, what they bought last year, their response times, which deals worked, and which deals failed.
The Intelligence Layer
This is the reasoning engine. It monitors the external world: news feeds and regulatory updates cross-referenced with buying and selling intent signals from socials, and filters it through the lens of the memory layer. Without this calibration, you get noise.

A bad model is like a CIA analyst with a neurodegenerative disorder — it sees everything but understands nothing.

A properly tuned intelligence layer flags a news item about “Private Equity acquiring secondary capabilities” and immediately identifies it as a specific strategic threat or opportunity, based on its knowledge of your business model.

According to Coalition Greenwich’s 2026 Market Trends, the disruption in finance will be AI digging into unstructured private markets data to find correlations humans miss. This validates our thesis: the alpha is in synthesizing the noise, not just summarizing it.

3. The Action Layer
Understanding without action is wasted compute. If the intelligence layer spots a market shift, the action layer automatically suggests: “Update the investor deck to reflect this new valuation metric.” If the system detects a liquidity crunch in a specific asset, it automatically drafts a tender offer for the relevant holders in our database.

It moves from “FYI” to “Execution,” ensuring that the distilled knowledge is applied to get tangible results.

4. Continuous Learning

This is the step most firms miss. If an agent writes a code snippet or drafts a matching email, the system must validate the outcome. If the benchmark improves (e.g., the code runs faster, or the client responds), the system updates its own “system prompts” to lock in that improvement. We recently saw an agent automatically expand a prompt from 200 tokens to 2,000 tokens because it learned that providing more context yielded a higher success rate. The loop coded that change itself.

This mirrors the “Service-as-Software” thesis seen in legal tech (Eve). As noted by a16z, Eve gets smarter at valuing lawsuits with every case it processes, creating a flywheel no competitor can catch. Similarly, every time our system processes a trade outcome, the shadow order book becomes sharper.

Understanding without action is wasted compute. The system moves from ‘FYI’ to execution, ensuring that the distilled knowledge is applied to get tangible results.

Converting Signals Into Actions

How does this apply to pricing assets? By building the loop to work with the communication channels like Telegram and Signal.

In private markets, liquidity often signals itself in informal text that is easy to miss. A message pops up in a group: “Looking to buy Arbitrum.” To a human, that’s a fleeting comment. To a spatially aware architecture, that is a data point for the shadow order book.

But raw text isn’t enough. We need validation.

The “Confidence Score” Validation
Trusting any message without a filter would fill an order book with junk. We solve the ‘Sybil’ problem of private markets by enforcing an identity check. The AI acts as the gatekeeper, filtering out the noise of window shoppers to reveal the signal of verified buyers.

The system analyzes intent using a validation logic loop:

Identify intent: Identify the asset and the action (Buy/Sell).
Check identity: Who is this user? The system queries the Memory Layer.
Validate background: Have they onboarded? Is their KYC complete? Have they traded before?

If the answers are “Yes,” the system assigns a high confidence score to that signal and converts a random chat message into a verified order inquiry.

Over time, this aggregation reveals the inquiry-based price floors that don’t exist on any public dashboard.

Moving From Reactive to Predictive Workflows

Currently, most brokerage workflows are reactive: “A client asked for X, let’s find it.”
The next phase is proactive: “This client holds Asset A; news just broke about Asset A; we should alert them.”

The end state is predictive.

If the system has enough data on a buyer, like their browsing history, past trades, and their “dwell time” on specific reports, it can predict liquidity needs before the client articulates them. It creates a match based on latent demand, ahead of time.

It’s based on algorithmic matching. If the system knows a buyer loves AI infrastructure and buys Series B rounds, it doesn’t wait for a request. When a relevant asset hits the shadow order book, the action is triggered, and a connection is made instantly. Previous intent based data is automatically converted into future actions, and these loops compound over time, increasing the personalized feel.

Beating Competitors With Data and Logic

In 2025, the AI model itself is not a moat. The difference between OpenAI, Anthropic, and open-source models is shrinking to a margin of less than 10%. They are commodities.

If you are relying on GPT-4 to be your edge, you have no edge.

The only defensibility left lies in two things:

Proprietary data. The transaction logs, status, behavior insights, and the granular history of who bought what. This is the fuel.
System prompts. The digitization of your unique strategic thinking: specific instructions, “do’s and don’ts”, verification protocols, and guardrails programmed into your agents.

As a16z’s David Haber notes, the only defensible strategy in the AI era is ‘owning the end-to-end workflow, becoming a system of record, becoming that system of action.’

For us, that record is the agentic shadow order book.

The Dawn of Autonomous Organization

We are moving toward a future where the maintenance and growth of the platform are driven by agents. The human role isn’t to process the trade or read the news; it’s not even verification, aka “human-in-the-loop”. The human role is to design the loop.

By building a system with spatial awareness, we capture the invisible liquidity that has slipped through the cracks before, and turn noise into a price.