30.31 Orchestrator-Workers

A central agent reads a complex task, decomposes it into subtasks at runtime, dispatches each to a specialist worker, and synthesizes the results. Unlike a Pipeline, the decomposition is generated by the model — not fixed at design time.


Motivating Scenario

A venture fund runs an AI agent for investment due diligence. Each new deal requires a different set of investigation workstreams: a fintech deal needs regulatory compliance analysis; a deeptech deal needs IP landscape research; a consumer deal needs cohort retention analysis. No two deals require exactly the same DD structure.

A fixed Pipeline cannot handle this — the stages differ by deal type. A human analyst reads the pitch deck and decides which workstreams to run. The Orchestrator-Workers pattern replaces this analyst: the Orchestrator reads the deck, generates a DD plan (which workstreams, in what order, what questions each must answer), dispatches specialist workers to each stream, and synthesizes results into an IC-ready memo. The workers — market researcher, team profiler, financial modeler, technical assessor — are fixed. What changes each time is which workers get activated and what questions they receive.

Structure

Zoom and pan enabled · Concrete example: VC due diligence orchestration

Key Metrics

MetricSignal
Decomposition coverage Do generated workstreams cover the full DD surface? Benchmark against expert-reviewed checklists by deal type.
Worker match accuracy Are workstreams dispatched to the correct specialist? Mismatch rate from registry validation.
Synthesis coherence score Cross-section consistency — how often does the IC flag contradictions in the memo?
Cost vs. fixed Pipeline Does dynamic decomposition generate enough quality improvement to justify the Orchestrator overhead?
NodeWhat it doesInputOutput
DD Orchestrator Reads pitch deck + fund thesis. Generates DD plan: which workstreams are required, what specific questions each worker must answer, what output format is expected. Dispatches workers in parallel or sequence based on dependencies. Pitch deck + fund thesis + worker registry DD plan: list of (worker, questions, context slice) tuples
Market Researcher TAM/SAM analysis, competitive landscape, market timing. Web search + database access. Market questions + company description Market section: size, competitors, timing signal
Team Profiler Founder backgrounds, prior exits, domain expertise, team gaps. LinkedIn, Crunchbase, news search. Founder names + roles Team section: credentials, red flags, network signal
Financial Modeler Unit economics, growth projections, burn rate, runway. Works from financials if provided. Financial data or estimates from pitch Financials section: unit econ, 3-year projection, key assumptions
Memo Synthesizer Combines worker outputs into IC-ready memo. Resolves contradictions across sections. Generates summary recommendation with confidence and key open questions. All worker outputs + original pitch deck Full IC memo: executive summary, sections, recommendation, open questions

When to Use

Use when
Avoid when

Value Profile

Origin of ValueWhere it appearsHow it is captured
Governance DD Orchestrator node The Orchestrator decides which workstreams exist, which questions each worker receives, and what counts as sufficient coverage. This is the governance act — it determines where value is generated across the entire DD process.
Future Cashflow Quality of decomposition A bad DD plan — wrong workstreams, wrong questions — loses value at the plan step, before any worker executes. The Orchestrator's judgment quality is the primary driver of overall output quality.
Value Transfer Orchestrator → Worker edges Workers capture value by executing assigned work and returning it to the Orchestrator. Their contribution is scoped — they transfer execution value back, not directly to the IC.
Representation Memo Synthesizer node The synthesized memo represents all worker contributions in a unified, coherent form. Its quality determines whether the sum of parts exceeds individual sections — synthesis is where representation value is created or destroyed.
VCM analog: Governance Token. The Orchestrator is the DAO allocating compute resources (workers) across tasks. Workers are contributors executing allocated work — their value is proportional to task completion, not orchestration. Adding a worker to the registry is analogous to a token holder joining a protocol.

Dynamics and Failure Modes

Orchestrator hallucinating workstream structure

For a healthcare deal, the Orchestrator generates a DD plan that includes "regulatory pathway analysis" as a workstream — but dispatches it to the Market Researcher, who has no regulatory expertise or data access. The Market Researcher produces a plausible-looking but fabricated regulatory analysis. The memo looks complete. Fix: Orchestrator must validate each workstream assignment against the worker registry — only dispatch to workers that explicitly support the requested workstream type. Worker registry entries must include capability declarations, not just names.

Context starvation in workers

The Financial Modeler receives: "Analyze unit economics. Company: Acme." It has no access to the actual financial data from the pitch deck — the Orchestrator scoped the context too narrowly. The modeler fabricates unit economics from industry benchmarks. The memo presents invented numbers as analysis. Fix: Orchestrator context scoping must be explicit. For each worker, trace which facts from the source document are required and include them in the dispatch. A context audit step before dispatch prevents this.

Synthesis incoherence

Market Researcher says TAM is $8B. Financial Modeler's revenue projection assumes a $1.2B addressable market. The Memo Synthesizer combines both sections without flagging the contradiction — it produces a coherent-looking memo with an internal inconsistency that a careful reader would catch immediately. Fix: Synthesizer must run a consistency check across sections before producing the final memo. Contradictions between sections must be flagged as open questions, not silently included.

Variants

VariantModificationWhen to use
Sequential Orchestration Workers execute in dependency order, not parallel; Orchestrator gates each dispatch Subtasks have hard dependencies — Team Profiler output needed before Financial Modeler can assess founder track record bonus
Adaptive Orchestration Orchestrator revises the DD plan after each worker returns results — adds or removes workstreams based on findings Task structure is only partially known upfront — e.g., a technical red flag triggers additional IP landscape workstream
Recursive Orchestration Workers are themselves Orchestrators for deeper subtasks Very complex tasks requiring multi-level decomposition — e.g., Market Researcher sub-orchestrates geographic market analysis

Related Patterns

PatternRelationship
10.11 PipelineStatic alternative — if workstreams are always the same across deals, Pipeline is simpler, cheaper, and more auditable
10.12 RouterSimpler routing — Router handles single-dispatch to one specialist; Orchestrator handles multi-dispatch to N specialists with a generated plan
30.31 Feedback LoopClose the loop — IC decisions on memo quality feed back to improve the Orchestrator's decomposition logic and Synthesizer's coherence

Investment Signal

The Orchestrator is the organizational IP. Workers are commodity — they can be replaced with better models from APIs. The Orchestrator encodes how the organization structures work: its decomposition logic, workstream taxonomy, worker registry, and context scoping rules represent accumulated institutional knowledge.

Integration = registering new worker capabilities in the Orchestrator's worker registry. Two firms using this pattern can merge at the Orchestrator layer by expanding the worker registry to include both firms' specialists — no architectural change required.

Red flag: an Orchestrator with no formal worker registry and no documented workstream taxonomy is opaque — it decomposes differently on every run with no reproducibility. Due diligence cannot determine what it actually does without extensive trace-log review. This is a valuation risk, not a technical curiosity.