Sequential chain of specialized agents. Each transforms the context packet before passing it forward. The routing logic is fixed at design time — no agent decides where to send work next.
A VC firm receives 300 deal memos per month. Each memo needs: (1) a one-paragraph summary for the IC, (2) a competitive landscape check against the portfolio, (3) a founder background pull from LinkedIn and Crunchbase, and (4) a go / pass / follow-up recommendation with reasoning. A single analyst doing this serially takes 45 minutes per memo. A pipeline of four specialized agents does it in 90 seconds.
The key insight: each subtask requires different capabilities and context. The summarizer needs compression skill. The competitive checker needs access to the portfolio database. The background puller needs web search. The recommender needs all prior outputs plus scoring criteria. These are not the same job — they are four specialists working in sequence.
| Metric | Signal |
|---|---|
| End-to-end accuracy | Primary quality signal — do IC recommendations match human analyst judgment? |
| Step-level error rate | Identifies which stage injects failures — Summarizer errors are the most expensive |
| Token cost per stage | ROI of each specialist — a stage where cost > downstream quality improvement is a removal candidate |
| Mean time to failure point | How deep into the pipeline errors typically propagate before detection |
| Node | What it does | What it receives | What it produces |
|---|---|---|---|
| Summarizer | Compresses raw memo into 200-word IC brief | Raw PDF / text memo | Structured summary: problem, solution, market, ask |
| Portfolio Checker | Checks for conflicts and overlap with existing investments | Summary + portfolio DB access | Conflict flag, similar portfolio companies, white-space assessment |
| Founder Researcher | Pulls founder backgrounds from LinkedIn, Crunchbase, news | Founder names from summary | Structured founder profile: exits, credentials, network |
| IC Recommender | Produces go / pass / follow-up with reasoning against fund thesis | Summary + portfolio check + founder profiles | Recommendation with confidence score and key questions |
| Origin of Value | Where it appears | How it is captured |
|---|---|---|
| Future Cashflow | Final output quality | Specialization compounds — each stage improves precision. Value realized at the Recommender node, not at intermediaries. |
| Governance | IC Recommender node | Recommender's scoring criteria encode the fund's investment thesis. This is the governance constraint over the entire pipeline output. |
| Conditional Action | Every node | Each stage requires compute to activate. Intermediate stages are pure cost unless their outputs are independently inspectable or reusable. |
| Risk Exposure | Summarizer (first node) | A bad summary at step 1 corrupts all downstream stages. Error amplifies forward — the pipeline has no backtrack mechanism. |
VCM analog: Work Token chain. Each agent earns its position by contributing to the final output. A stage that does not measurably improve quality is removable without loss — the same test applied to underperforming nodes in a Work Token system.
A bad output at the Summarizer corrupts Portfolio Checker, Founder Researcher, and IC Recommender — in that order. The pipeline completes and delivers a confidently wrong recommendation. Detection requires per-step evaluation, not just end-to-end scoring. Fix: validate Summarizer output against a schema before passing forward, or add a lightweight quality gate after step 1.
Each agent sees only the previous agent's output. The Founder Researcher never sees the original memo — only the summary. If the Summarizer dropped a critical detail (e.g., that the founder previously worked at a portfolio company), the Founder Researcher has no way to recover it. Fix: pass original context alongside artifacts at each step, or use a Context Funnel pattern upstream that preserves key fields.
The pipeline assumes Portfolio Checker can run before Founder Researcher. If portfolio conflict detection requires knowing the founders first (e.g., to check co-investor network), the ordering is wrong and the pipeline produces invalid results silently. Fix: model stage dependencies explicitly before fixing the order.
| Variant | Modification | When to use |
|---|---|---|
| Pipeline with Retry | Final validation node loops back to previous stage once on failure | Quality gate matters; single retry is sufficient (e.g., IC Recommender rejects and Founder Researcher re-runs with expanded query) |
| Context-Preserving Pipeline | Original input passed alongside artifacts at every step | Downstream agents need access to source material, not only prior agent output |
| Branching Pipeline | One stage fans out to two specialized sub-pipelines, merges before output | Two parallel tracks (e.g., market analysis + team analysis) that converge at the Recommender |
| Pattern | Relationship |
|---|---|
| 10.15 Evaluator-Optimizer | Add a feedback loop at the end — Recommender rejects and loops back to Researcher for more data |
| 20.23 Orchestrator-Workers | When stage order cannot be fixed at design time, replace with a dynamic Orchestrator |
| 30.31 Feedback Loop | Close the loop — IC decisions feed back to update the Recommender's scoring criteria over time |
Pipelines are the most legible architecture for due diligence. Every stage is a measurement point — acquirers can audit exactly where value is added or lost. A well-instrumented pipeline exposes its own health: stage-level accuracy, cost, and latency are all independently auditable.
Integration = splicing two pipelines at a compatible interface. If firm A's Researcher output schema matches firm B's Recommender input schema, the merger is an architectural operation, not a cultural one.
Red flag: a pipeline with no per-stage instrumentation cannot be audited. If the firm only measures final recommendation quality, step-level value attribution is impossible — and so is pricing the acquisition correctly.