Chapter 8

Specialized agents keep each context window clean

When executive leadership envisions AI deployment, they often imagine a single autonomous analyst.

They imagine an executive assistant model that can ingest a three-hundred-page regulatory filing, extract financial statements, calculate valuation multiples, perform cross-industry benchmarking, write an investment memorandum, and audit its own text for compliance.

In the real world, asking a single prompt to perform an entire end-to-end research workflow guarantees operational failure.

When one prompt attempts to manage extraction, arithmetic, drafting, and compliance checks simultaneously, its context window rapidly clutters. As instructions compete with raw data, attention degrades, negative constraints drift, and early hallucinations get swallowed as ground truth.

To achieve reliable institutional execution, enterprises do not rely on omniscient universal prompts. They organize work across specialized agent pipelines.

Clean context handoffs prevent error compounding

What is an AI agent?

In modern software architecture, an agent is not a sentient digital worker. An agent is simply a language model invoked inside a loop with a specific system role, access to specific tools, and a tightly constrained context window.

The fundamental rule of institutional pipelines is context isolation:

Instead of forcing one model to hold the entire project history, you decompose the workflow into distinct roles. Each role executes its sub-task inside a pristine context window, writes a structured output, and terminates.

Close

Consider an equity research pipeline:

1. The Extractor Role: Given only the raw 10-K filing, this agent has one mission: locate Note 8 and extract credit facility terms into clean JSON. It writes zero narrative commentary and performs zero math.
2. The Quantitative Role: This agent receives the extracted numbers and executes external calculator tools to compute net leverage and interest coverage. It has no access to the prose documents; its context window contains only mathematical schemas.
3. The Red-Team Auditor: This agent compares the extracted numbers against original document hashes and verifies that footnote caveats were preserved. If an extraction fails verification, it rejects the package back to the extractor.
4. The Drafting Editor: Once all data is validated, the editor receives the audited facts and synthesizes the executive memorandum for human review.

Because each agent operates over a small, targeted context window, attention remains razor-sharp. Errors are contained at the boundary of each step rather than propagating through an unmanageable sixty-turn conversation.

Digital segregation of duties stops runaway consensus

In institutional governance, one of the oldest risk management principles is segregation of duties (SoD).

Under statutory internal control frameworks, an organization never permits the same individual to authorize a financial transaction, record the journal entry, and reconcile the bank account. Concentrating authority creates opportunities for fraud and unspotted mistakes.

The same principle applies directly to artificial intelligence:

If you allow the same agent that extracted an accounting figure to audit its own extraction, it will almost always certify its own work. Large language models exhibit strong confirmation bias within a single context window. Once a model commits to a claim, it interprets follow-up prompts to justify its previous output.

In multi-agent pipelines, institutions enforce digital segregation of duties:

The extraction agent cannot serve as the verification agent. Furthermore, critical handoffs must incorporate deterministic code gates: hardened validation programs that enforce schema checks before passing data to the next agent. Finally, every pipeline must terminate at an air-gapped human sign-off checkpoint before publication.

Document:
[Excerpt from Acme Industrial Technologies 10-K, Note 8: Debt Covenants and Credit Facilities:
"Under the Senior Credit Facility, consolidated funded debt to EBITDA shall not exceed 3.50 to 1.00 at the close of any fiscal quarter. Consolidated funded debt stood at $2,800 million against cash and equivalents of $360 million. Consolidated EBITDA for the trailing twelve months reached $1,220 million. Consolidated fourth-quarter revenue totaled $4,280 million."]

Role: Document Extractor
Instructions:
Extract the credit metrics into a strict JSON object.
Do not write narrative commentary or perform mathematical calculations.
Use keys: company, fourth_quarter_revenue_millions, funded_debt_millions, cash_millions, trailing_ebitda_millions, covenant_ratio_limit, and covenant_source.

{
  "company": "Acme Industrial Technologies",
  "fourth_quarter_revenue_millions": 4280,
  "funded_debt_millions": 2800,
  "cash_millions": 360,
  "trailing_ebitda_millions": 1220,
  "covenant_ratio_limit": 3.50,
  "covenant_source": "10-K Note 8"
}