SecondLane's Agent Stack for Private Market Matchmaking

SecondLane, a nine-person secondary market advisory firm based near New York, runs a 16,000-party CRM and a business intelligence pipeline processing roughly 6,000 signals per day at a total cost of approximately $600 per month across two companies. The system runs on Anthropic Claude, Pipedream, and Obsidian. Private secondary markets require constant counterparty enrichment, real-time regulatory monitoring, and rapid document generation, all of which previously depended on expensive specialized human labor.

The firm operates in an advisory and matchmaking capacity, connecting buyers and sellers of digital assets, SAFTs, equity, and pre-IPO company shares. As of March 2026, SecondLane is four months into an alternative trading system and broker-dealer license application. The licensed execution platform is in development pending those approvals.

Business Context

Forge Global, the largest pre-IPO secondary market platform before its acquisition by Charles Schwab for approximately $660M in March 2026, operated at a structural loss throughout its public life2. Public financials show that labor compensation alone exceeded total revenue in recent quarters. The compliance infrastructure to support that labor added further cost on top of that.

The economics reflect a wider pattern in secondary markets: highly specialized deal professionals, extensive legal and compliance coverage, and continuous counterparty relationship maintenance all scale poorly with headcount. SecondLane's operating thesis is that the same work can run on digital labor. The firm reported over $100M in transacted volume and $100M in assets under management for 2024 with a team of approximately nine people2.

The AI operating layer is the enabling constraint. It lets the team match, enrich, and communicate at a volume that would require a far larger staff under conventional operations.

The crypto-native side of the business runs primarily through Telegram, where most counterparties originate and communicate. Traditional finance counterparties cannot use Telegram under their own regulatory constraints, so the system maintains a parallel channel for that segment. This creates the core architectural requirement: a single orchestration layer that pushes the same enriched counterparty data through different delivery channels without human intervention per transaction.

Architecture

SecondLane structures its operating system around four layers: memory, intelligence, action, and a learning loop. The memory layer is built on Obsidian with a markdown file per entity. Those files form the base for relational queries and feed the knowledge graph.

The intelligence layer runs continuous signal processing from external sources and applies it back to the firm's own stack and counterparty profiles. The action layer executes document generation, CRM enrichment, and outbound communication through pre-defined agentic workflows. The learning loop feeds observability data back into the system via Langfuse traces.

graph TD
    A[External Signal\ne.g. arXiv paper\nSEC filing\nmarket price update] --> B[Relevance Filter\nSignal scored 0-10\nagainst company goals]
    B -->|Score >= 8| C[Instant Notification\nto operator]
    B -->|Score < 8| D[Queued Summary\nor Discarded]
    C --> E{Human Go/No-Go}
    E -->|Go| F[Agent Dispatched\nbenchmark or action]
    E -->|No-Go| G[Dismissed]
    F --> H[Langfuse Trace\nLLM calls, tool calls\nagent runs, user sessions]
    H --> I[Learning Loop\nsystem self-improvement]

    J[Inbound Order\ne.g. Anthropic shares] --> K[CRM Enrichment\nObsidian + Apache AGE\n16,000 party database]
    K --> L[Counterparty Match\nIntent scoring\nhistory lookup]
    L --> M{Human Review}
    M -->|Approved| N[Outbound via\nTelegram / Email / Slack]
    M -->|Revise| K

Pipedream acts as the integration bus, connecting Claude, Exa, Telegram, Google Drive, and Google Calendar to distribution partners. All workflows are pre-designated: the team defines the permitted action graph before execution, not at runtime.

Agent permissions are enforced at the role level. Each agent is configured with an explicit list of permitted actions. No agent can take an action outside its approved set without human approval.

This eliminates a class of failure where agents act on partial information across systems they should not touch.

The Head of Strategy and Head of AI owns the initial permission configuration. The CTO maintains it on the engineering side.

Stack

SecondLane previously used Lindy.ai and Zapier for workflow automation. Both were replaced because they require humans to explicitly define every process. The team concluded that a Claude Code-based approach reduces the cycle time for adding or modifying workflows from days to hours.

The system has been rebuilt three times. Each rebuild was treated as a module-level refactor rather than a full rewrite. The team isolated the components that improved and replaced only the pieces that the new architecture handled better.

Outcomes

SecondLane's BI pipeline processes approximately 250 signals per hour across regulatory filings, arXiv preprints, market pricing data, and competitor activity. The system runs across two companies at a combined cost of approximately $600 per month.

The system's signal scoring is customized to SecondLane's actual stack and investment thesis rather than producing generic market summaries. Anything scoring 8 out of 10 or higher on relevance triggers an instant operator notification. Below that threshold, signals are queued or discarded.

The CRM holds approximately 16,000 enriched counterparty records. When an order arrives, the system identifies the highest-probability counterparty matches from the database without human search.

A test integration with Republic aggregated approximately $25M in buy-side interest across 10,000 unique parties. The orders fell below SecondLane's typical minimum, but the intent data informed a product direction: tokenizing assets into a special purpose vehicle could aggregate smaller orders into tradeable blocks.

For 2024, SecondLane self-reported over $100M in transacted volume and $100M in AUM with a team of nine.

Failure Modes

Permission misconfiguration during early development. The team experienced agent behavior that deleted or corrupted data in early testing. The root cause was unrestricted agent capability: a model with access to a data store and no action boundary will take destructive actions under plausible instructions.

SecondLane's response was a hard constraint: no agent may take an action outside its pre-approved set. High-consequence actions require explicit human approval before execution.

Split channel architecture adds operational complexity. Crypto-native counterparties use Telegram. Many traditional finance counterparties are prohibited from using it, so the system maintains two parallel delivery paths for otherwise identical enriched data.

Pre-designated workflows cannot self-extend. SecondLane's agent system uses pre-defined workflow graphs. The team has not deployed automated workflow capture. That feature is under active development at Chiri, a separate company co-founded by SecondLane leadership. It is not yet in production at SecondLane2.

Knowledge graph is newly deployed and untested at scale. Apache AGE was introduced the day before the interview. The team does not yet have production data on how it performs under the full 16,000-party load2.

Regulatory status limits automated execution. The platform cannot operate as a licensed broker-dealer until the ATS license is granted. All agent workflows today stop at document generation, CRM matching, and communication drafts. A human must approve before anything goes to a counterparty2.

Lessons

The core architectural decision at SecondLane is to treat every agent action as an auditable event from day one. This was not purely a compliance choice. The Langfuse trace that satisfies SEC audit requirements is the same trace that feeds the learning loop2.

Document generation is the highest-return starting point for a brokerage or advisory workflow. NDAs, service agreements, and referral agreements are templated, predictable, and high-frequency. Starting there gives a team a working production loop with real volume before touching harder communication or execution problems.

Obsidian as a persistence layer is not a common enterprise choice, but the markdown-per-entity model has a specific advantage for agent workflows: every file is machine-readable, version-controlled in Git, and writable by any tool that can edit text.

The rebuild count matters less than the module structure. SecondLane rebuilt its system three times in short iterations because the components were modular. Teams that build monolithic agent pipelines will face the same pace of model and tool improvement without the ability to swap components independently.

Related guide: Open the current guides build