Portfolio Listing Agent and Current Portfolio Section

Design and behavior of the portfolio_listing_agent that exhaustively enumerates current portfolio companies and wires them into the Portfolio Construction section for LP-commit emerging manager memos

Problem

For LP-commit / emerging manager fund memos, the Portfolio Construction section needs to:

• Explicitly enumerate all current portfolio companies that appear in the deck or dataroom
• Avoid hallucinating names or funds that are not actually listed in those materials
• Provide short, consistent descriptions and links for each company

Historically, this responsibility was left to generic section writers (research + writer agents) that:

• Often mentioned one example company (“representative investment”) but did not list the full portfolio
• Were prone to hallucinating company names based on sector themes
• Produced inconsistent detail (some companies deeply described, others omitted)

We needed a dedicated agent that treats the portfolio list as a first-class artifact and a source of truth.

High-Level Design

Goals

• Exhaustive enumeration of all portfolio companies visible in internal materials (deck, state.json, dataroom-derived data)
• Structured representation saved as current_portfolio.json
• Deterministic markdown subsection ## Current Portfolio injected into the Portfolio Construction section (04-portfolio-construction.md)
• Link and logo enrichment that stays grounded in already-listed companies

Scope

The portfolio_listing_agent is only intended for:

• Fund memos (investment_type == “fund”)
• Emerging manager LP commit outline: outline: "lpcommit-emerging-manager" in the company JSON

It should be no-op for:

• Direct company memos
• Other fund memo outlines that do not need exhaustive portfolio enumeration

When the Agent Runs

Trigger Condition

The agent is integrated into the enhancement workflow for fund memos and gated on the outline:

if state.get("outline_name") == "lpcommit-emerging-manager":
    portfolio_listing_agent(state)

Notes:

• outline_name is populated from the company JSON field outline (e.g. "lpcommit-emerging-manager").
• If outline_name is missing, we can safely derive it from state["company_data"]["outline"] when constructing MemoState.

Position in Workflow

The intended ordering is:

1. Deck / dataroom analysis (if present) populates deck_analysis and any initial section drafts.
2. Research agent enriches state["research"].
3. Portfolio Listing Agent reads state, the latest output directory, and constructs:
   • current_portfolio.json
   • ## Current Portfolio subsection under Portfolio Construction
   • Calls link_enrichment_agent to add hyperlinks to portfolio companies
4. Writer agent can still refine the Portfolio Construction narrative, but the portfolio list is already present and should be treated as source-of-truth.

This ensures that:

• The agent has maximum context (deck, state.json, research.json) before running.
• The writer agent doesnt have to guess the portfolio list; it can refer to the structured artifact.

Inputs and Data Sources

The agent reads from the same artifact stack used by other agents:

• Company name and metadata from MemoState / data/<Company>.json
• State file: output/<Company>-vX.Y.Z/state.json
  • Especially deck_analysis and any portfolio-like fields that already exist
• Research file: output/<Company>-vX.Y.Z/1-research.json
• Existing sections: output/<Company>-vX.Y.Z/2-sections/*.md

Within these, it looks for:

• Slide- or state-level references to specific portfolio companies
• Any already-structured portfolio data (e.g., lists of investments in deck analysis)
• Mentions of portfolio holdings in the existing Portfolio Construction section

The rule is: only list companies that appear in the deck / internal materials or derived structured data, never purely from web research.

Outputs

1. current_portfolio.json

A structured artifact in the output version directory, with a schema along the lines of:

{
  "fund_name": "Watershed VC",
  "as_of": "2025-04-03",
  "portfolio_companies": [
    {
      "name": "Watershed Health",
      "description": "Seed-stage health data platform for value-based care",
      "thematic_fit": "HealthCare / care delivery transformation",
      "stage": "Seed",
      "check_size": null,
      "initial_investment_date": null,
      "company_url": "https://...",
      "portfolio_page_url": "https://fund-site/portfolio/...",
      "logo_url": "https://.../logo.png",
      "notes": "Derived from deck; missing explicit round size."
    }
  ]
}

Design principles:

• Prefer explicit values from the deck; mark unknowns as null or "Unknown" rather than guessing.
• Include a short 50 6 character description for each company, tuned for table/list usage in memos.
• Include URLs where possible:
  • Company site
  • Fund portfolio page entry
  • Logo URL (if discoverable)

2. Markdown Subsection: ## Current Portfolio

The agent writes or updates a subsection inside 2-sections/04-portfolio-construction.md:

• Heading: ## Current Portfolio
• Body: markdown list or table of all portfolio_companies from current_portfolio.json
• Each row includes:
  • Company name (with markdown link when available)
  • Short description (50 6 characters)
  • Stage and thematic fit (if known)

This subsection is designed to satisfy the outlines validation criteria for Portfolio Construction:

• EXHAUSTIVE list of all current portfolio companies in the fund (as seen in deck/dataroom)
• Per-company description and link

3. Link Enrichment

After the markdown subsection is created/updated, portfolio_listing_agent invokes the existing link_enrichment_agent to:

• Add hyperlinks to the first mention of each portfolio company
• Prefer official company sites, fund portfolio pages, or authoritative profiles

Crucially, link enrichment is allowed to add URLs, but not new company names.

Link and Logo Discovery Strategy

The agent follows a conservative enrichment strategy:

1. Fund website portfolio pages
   • Use company_data["url"] (e.g., https://www.watershed.vc/) as the base
   • Look for /portfolio, /companies, /investments, /funds paths
2. Per-company portfolio entries
   • From the fund site, identify individual company pages matching names from the deck
3. Company websites and logos
   • For each known company name, search for:
     • Official website
     • Logo asset URLs (png/svg/jpg)

If the agent cannot confidently match a URL to a company, it leaves the URL fields blank and does not invent.

Safety and Anti-Hallucination Rules

To keep the portfolio list trustworthy:

• No synthetic companies
  • A company must be explicitly mentioned in the deck, dataroom-derived JSON, or existing internal state.
• No synthetic allocations or outcomes
  • Do not infer check sizes, ownership, or realized returns unless those numbers are explicitly present.
• No overfitting to web research
  • Web research is allowed to fill in missing URLs and logo paths, not to invent portfolio names.
• Describe, dont overclaim
  • Use short, factual descriptors (“Seed-stage cardiac monitoring platform”), not marketing language.

If input signals are weak or inconsistent, the agent should still:

• Produce a minimal but explicit ## Current Portfolio subsection, possibly with notes that certain details are missing
• Avoid filling gaps with speculation

Failure Modes and Degradation Behavior

1. Missing Output Directory / Artifacts

If the agent cannot find an existing output directory for the company (e.g., early in a pipeline, or an unexpected path issue):

• Log / print a message like: "⊘ Portfolio listing skipped - no output directory found"
• Return without raising, so the rest of the workflow continues

2. No Deck / No Portfolio Signals

If there is no deck, state.json, or research.json with identifiable portfolio companies:

• Write a stub current_portfolio.json with an empty portfolio_companies list
• Optionally add a short note in the ## Current Portfolio subsection indicating that no current portfolio companies were identified in available materials
• Do not attempt to infer portfolio holdings from public web research alone

3. API / Network Failures for Enrichment

If link/logo enrichment fails (Anthropic/Perplexity/network issues):

• Leave current_portfolio.json in place with whatever structured data exists
• Keep the ## Current Portfolio subsection without enriched links
• The memo remains usable; the portfolio list just has fewer conveniences

Interaction with Outline and Validation

For lpcommit-emerging-manager outline, the outline file encodes:

• That Portfolio Construction must include an exhaustive list of portfolio companies
• That each company must have short description + link + context (stage, thematic fit)

The combination of:

• portfolio_listing_agent (structured enumeration and markdown) and
• Outline validation_criteria (content contract)

is intended to:

• Make the portfolio list deterministic and repeatable across runs
• Reduce the chance that later agents or edits accidentally drop companies or introduce hallucinated holdings

Future Extensions

• Integrate more tightly with a Portfolio Data MCP server to:
  • Cross-check deck-derived portfolio against firms source-of-truth portfolio database
  • Flag discrepancies (e.g., deck mentions a company not in portfolio DB)
• Add allocation-aware views (e.g., % of capital by sector, geography), while still refusing to guess missing allocations
• Support multi-fund complexes, where one GP runs several vehicles and the memo must clarify which companies sit in which fund.