AgenticOS

Five agents · one pod · institutional finance

Enterprise Financial Agentic OS

The arithmetic leaves the model. The decision stays with the human.

Five system-prompt packages that organise an AI pod the way a firm is organised — intelligence, production, governance, coordination, orchestration. Each one deploys into ChatGPT, Gemini, Claude or Grok in about a minute, and none of them can send anything to a client on their own.

Agents05
Files17
On disk74.1 KB
Platforms04
Meet the pod Download all five Agent Library

The operating model

The argument is not that fewer people are needed. It is that the shape is wrong. A pyramid spends its widest layer on the work least worth a human — so the pod absorbs that layer and the pyramid becomes a diamond.

Traditional pyramid

Widest where it matters least

Senior leadershipJudgement, relationshipsnarrow
Mid-level VPsTranslation, checkingmiddle
Junior analystsExtraction, formatting, rebuilding links60%

Augmented diamond

Widest where judgement compounds

Chairman / human MDDirection, and the only signature35%
Agent 5 · Chief of StaffRouting, mediation, budget50%
Agents 1–4 · the podData, modelling, compliance, strategy15%

The pod

Five packages, each a system prompt plus a grounding knowledge base. They are written to be read by a person before they are pasted into a model — the constraints are the interesting part, not the capabilities.

Data & Intelligence

The Autonomous Prospector

Turns unstructured market noise into structured, cited, machine-readable fact.

The firm’s continuous intelligence engine. It absorbs gathering, cleaning, structuring and monitoring so analysts spend their attention on conviction rather than collection.

Database Interrogator

Conversational request → optimised, read-only SQL. Schema introspection over MCP before execution; destructive statements (DROP, DELETE, UPDATE, ALTER) are blocked outright.

Speed Reader

Metadata-tagged chunking across 10-K, 10-Q, 8-K, S-1 and call transcripts. Zero-shot risk classification and tone-shift detection between sequential quarters.

Alpha Scraper

Agentic ETL that adapts to layout changes instead of breaking on them. Hiring velocity, pricing indices and review sentiment as leading signals.

Quant Translator

FinBERT-domain sentiment, so “churn is up a lot” reads as negative. Entity resolution maps Meta / Facebook / FB to one CIK.

Architecture

  • Heartbeat and cron loops — an always-on filing radar, not a chatbot
  • Swarm fan-out: SEC reader, DB interrogator and scraper run in parallel, then synthesise
  • Semantic + procedural memory so context is never re-explained
  • KNOWLEDGE_BASE.md2.7 KB
  • PLATFORM_IMPORT_GUIDE.md1.5 KB
  • PROMPT.md8.0 KB
Temperature0.1 Top_p0.9 Package12.1 KB
PROMPT.md KNOWLEDGE_BASE.md

Execution & Modelling

The Production Engine

Writes the formula. Never guesses the arithmetic.

The production co-pilot. It builds the model and the book so associates argue about the assumption rather than the broken link at 2 AM.

Deterministic execution

Language models are unreliable arithmetic engines, so the arithmetic leaves the model. Formulas are authored here; a Python or Excel engine behind MCP computes the result.

Model construction

Multi-statement models, 40-tab LBOs, synergy waterfalls. Conversational scenarios — “downside, 15% revenue drop, 50 bps on the cost of debt” — recompute the capital stack in place.

Book and CIM drafting

Comps refresh dynamically; Excel-to-deck links update rather than being pasted. House typography and spacing applied on generation.

Parallel assembly

A 100-page CIM is decomposed and drafted concurrently — industry overview, revenue build and management biographies — then integration-reviewed for consistency.

Architecture

  • Plan-Execute rather than open-ended ReAct — the plan is a contract, signed before work
  • Procedural memory encodes a specific MD’s formatting rules and applies them unasked
  • Idempotent tool schemas, so a retried call cannot corrupt file or database state
  • KNOWLEDGE_BASE.md3.0 KB
  • PLATFORM_IMPORT_GUIDE.md1.4 KB
  • PROMPT.md7.6 KB
Temperature0.1 Top_p0.9 Package11.9 KB
PROMPT.md KNOWLEDGE_BASE.md

Governance & Review

The Compliance Overlord

Generates nothing. Approves nothing quietly.

The auditor. It reviews, critiques and validates every artifact the other agents produce, and it is the only agent in the pod that can stop the line.

Generator–critic loop

Intercepts drafts and returns a specific, actionable verbal critique — which step failed, which figure was wrong, which filing and line to recompute from — then routes the task back.

Constitutional guardrails

Screens for MNPI exposure (SEC Rule 204A, FCA MAR), forward-looking guarantees and Chinese Wall breaches before anything reaches a client. A detection halts dispatch.

Human-in-the-loop gates

Non-bypassable. Execution pauses and packages an executive summary, verified citations and an explicit approve / reject decision for a named human.

Immutable audit log

Time-stamped record of every tool call, parameter and validation step — so a challenged multiple can be traced to the filing, page and line it came from.

Architecture

  • Verification loop scored against an explicit rubric, retried until it clears
  • Hard iteration cap of five cycles, then the loop is killed and escalated
  • Separation of concerns — the auditor never grades its own generation
  • KNOWLEDGE_BASE.md2.8 KB
  • PLATFORM_IMPORT_GUIDE.md1.4 KB
  • PROMPT.md8.1 KB
Temperature0.0 Top_p1.0 Package12.3 KB
PROMPT.md KNOWLEDGE_BASE.md

Coordination & Strategy

The Digital Vice President

Takes an objective in a sentence. Returns a decision in a dashboard.

The single point of contact for a managing director. It decomposes the goal, dispatches the pod, reroutes failures and frames the result as an argument rather than a spreadsheet.

Hierarchical decomposition

A stated objective becomes staged delegation — ingestion, modelling, audit, framing — with dependencies made explicit rather than assumed.

Dynamic rerouting

A missing filing or a dead URL is a reroute, not a crash. The task moves to an alternative source over MCP and the thread continues.

Narrative framing

Applies the tone a specific audience needs — how downside is framed for an investment committee is not how it is framed for a founder-CEO.

Preference retention

Corrected once, applied thereafter. Slide order, leverage bias and formatting become procedural rules on every future deal for that MD.

Architecture

  • Master loop controller across multi-week deal windows, resisting context rot
  • Persistent checkpoints — a project paused for a week resumes with full context
  • Engineered never to press the final button: the decision dashboard is the terminus
  • KNOWLEDGE_BASE.md2.2 KB
  • PLATFORM_IMPORT_GUIDE.md1.4 KB
  • PROMPT.md8.1 KB
Temperature0.2 Top_p0.95 Package11.7 KB
PROMPT.md KNOWLEDGE_BASE.md

Orchestration & Meta-optimisation

The Chief of Staff

Organises the other four into a firm, not a crowd.

The central nervous system. It routes by state, mediates when specialists disagree, holds the memory of past deals and enforces the budget.

Four-state routing engine

Information gap → Prospector. Production required → Production Engine. Strategic framing → Digital VP. Risk or final review → Compliance. The state, not a human, picks the agent.

Debate and mediation

When strategy proposes leverage that compliance flags as a coverage risk, it does not pick a side — it commissions a revised model and presents both to the human.

Three-depth memory

Semantic (firm guidelines), episodic (what went wrong on the last deal of this shape) and procedural (how this MD works).

Cost guardrails

Routine extraction goes to small models; frontier models are reserved for reasoning that needs them. No-progress detection terminates loops before they bill.

Architecture

  • Decompose–Plan–Merge over a dependency tree, with persistent state
  • Collaboration topology switches per task: sequential, parallel swarm, or critique loop
  • Meta-optimisation — execution logs are read back into the checklists
  • KNOWLEDGE_BASE.md2.8 KB
  • PLATFORM_IMPORT_GUIDE.md1.4 KB
  • PROMPT.md7.9 KB
Temperature0.2 Top_p0.95 Package12.1 KB
PROMPT.md KNOWLEDGE_BASE.md

A deal, end to end

One directive — assess a buyout and prepare the board pitch — moving through the pod. The state decides which agent runs, not a human clicking through a queue. It terminates at a person, by design.

  1. 01

    Information gap

    agent 1 · prospectorParallel swarm

    Ingests the filings, runs the SQL, pulls the alternative signal.

    Revenue and EBITDA extracted with page-and-line citations.

  2. 02

    Production required

    agent 2 · production engineSequential

    Builds the model against a signed plan; arithmetic executes deterministically.

    Returns, sensitivity grid and a first-draft book.

  3. 03

    Strategic framing

    agent 4 · digital vpSequential

    Frames the result for the audience that has to decide on it.

    Memorandum and client correspondence, drafted not sent.

  4. 04

    Risk and final review

    agent 3 · complianceReflexion loop

    Recomputes the arithmetic, checks every citation, screens for MNPI.

    Cleared, or returned with a specific critique.

  5. 05

    Governance checkpoint

    agent 5 · chief of staffEscalation

    Merges the outputs and stops at the human.

    A decision dashboard. Nothing dispatches without a signature.

Deploy

Nothing to install. Each agent is a prompt and a knowledge file; the platform supplies the model. Set the temperature listed on the agent’s card — the compliance agent is deliberately at 0.0.

  1. Create a GPT, open Configure.
  2. Paste PROMPT.md into Instructions.
  3. Upload KNOWLEDGE_BASE.md under Knowledge.
  4. Set temperature to the value on the agent’s card.

The file

What is actually in the download, and what this is not. Every figure below is read off disk by the build script that generates this page, so the page cannot drift from the artifacts.

What each package contains

  • PROMPT.md — the full system prompt: role, capabilities, architecture, guardrails and a worked output format.
  • KNOWLEDGE_BASE.md — the grounding document: lookup tables, schemas, rubrics and routing matrices the prompt refers to.
  • PLATFORM_IMPORT_GUIDE.md — click-by-click deployment for each of the four platforms.

Provenance of the two figures

The architecture cites a 74% token reduction for company-shaped agent hierarchies over flat ones, and up to 60% cost reduction from routing routine extraction to smaller models. Both come from the OrgAgent line of research and from published model-routing benchmarks — they are reported results from the literature, not measurements taken on this pack. Treat them as the design rationale they are. Nothing here has been benchmarked in production.

Measured on disk

Agent packages05
Markdown files17
Total bytes75,882
Total size74.1 KB
Target platforms04

Notice

  1. This is a prompt and knowledge architecture, not software. There is no runtime, no API, no deployment. It configures models you already have.
  2. Every figure in the worked examples — the target company, the returns, the audit log — is illustrative. Apex Technologies is not a real issuer.
  3. The compliance agent screens for MNPI and regulatory exposure as a design pattern. It is not a substitute for your firm’s compliance function, and it has not been reviewed by one.
  4. The human approval gate is written as non-bypassable in the prompt. A prompt is not an access control. If the decision matters, enforce it in your systems too.
  5. Released under MIT. Use it, fork it, argue with it. Attribution appreciated, not required.

System prompt

Loading…
Download