Comps agent
Screens peers through your data connectors, formats the table, and surfaces raw multiples. You set the peer set, because true comparability is judgment the model cannot fake.
Armory · Agent Packs
Pipelines that do the grunt work, and flag their own blind spots.
n8n and Claude agent workflows for comps, precedents, and IC memos, built around full-workbook comprehension and wired with human checkpoints where judgment is non-negotiable.
Why it matters
A good prompt saves minutes. A pack saves the whole evening.
An agent pack strings the mechanical steps, screen the peers, pull the filings, extract the figures, format the table, draft the memo, into one supervised pipeline, so you stop copying numbers between tabs and start deciding whether the numbers are right. Up to roughly 73 percent of foundational analyst hours on tasks like these can be automated.
Agents fail in specific, knowable ways: they hallucinate figures, break on circular references, misapply GAAP and IFRS, and lose context between sessions. So every pack is checkpointed by design. The pipeline is fast precisely because it knows where it is not allowed to be trusted.
Source: Knowledge / AI Capability Assessment 2026
/Hours back, structurally.
Multi-step pipelines compress screening, extraction, and first-draft memos into one run, on top of the 3 to 5 hours a week scenario automation already saves.
/In control at every gate.
Agents draft; you approve at each checkpoint. That is the pattern large banks already run, not a leap of faith.
/Deploy, not theorize.
Fork the workflow, plug in your tools, run it. These are working pipelines, not slide-ware.
/Keep the judgment that pays.
Synergies, capital-structure nuance, and deal narrative stay human-required. The pack hands those decisions back on purpose.
What's inside
Forkable workflows with checkpoints and failure modes on the tin.
Precedents agent
Surfaces comparable transactions, computes EV/EBITDA, and flags deal-specific gaps. Eight man-weeks compressed to hours, but the agent flags rather than concludes.
Memo agent
Structures an IC-memo first draft from your notes and data. You own the thesis and the risk weighting, full stop.
Guardrails
Every pack ships with its failure modes named and a human-checkpoint map showing exactly where you review before output moves forward.
Proof and credibility
The backbone is full-workbook comprehension: Claude reads every tab, formula, and dependency before it acts, which lets a pipeline operate on a real model instead of a flattened export. The same assessment is candid about the limits, which is why every pack forces a human gate before output. We build the speed and we build the catch.
Source: Knowledge / AI Capability Assessment 2026 · Perplexity research, 2026
Straight answers
Before you ask.
Isn't agentic AI overkill for me?
It is not all-or-nothing. Even AI around the edges saves 2 to 4 hours a week. Start with one agent on one painful step and scale into full pipelines when ready.
Will it run on my stack?
The packs are built on n8n and Claude and designed to be forked and adapted to your environment. Each pack documents the connectors it expects.
Is it compliant?
Compliance is your firm's responsibility. What the packs do is document the data-handling and audit-trail gaps so you can layer the right controls around them. Eyes open, not blind trust.
Will this hollow out my skills?
No. Because you review and approve at every gate, you cannot run a pack on autopilot. It accelerates a skilled operator; it does not manufacture one.
