Start with real documents, not training data
Parse your actual return, your actual contracts, your actual medical records. The grounding specificity is the entire point.
The Tax Strategist and Finance Agents
Tier 3 · Real Build
6 min read
The tax strategist and finance agents are the same architecture as the health-expert and professor agents, applied to a different domain. KB-backed, Opus-tier, grounded in real personal documents. What makes them distinct is the source material: the KB is not a scraped research corpus. It is JD’s actual 2024 tax return and 2025 W-2 PDFs.
The tax strategist
Section titled “The tax strategist”agents/tax_strategist shipped via the tax-strategist-2025 project (CHANGELOG entry). The agent is backed by parsed 2024 return + 2025 W-2 PDFs and a preparer evaluation document.
The first test query immediately surfaced a concrete number: the BAA HANE K-1 was identified as the highest-leverage tax move, with an estimated ~$14,300 swing. That is a specific, verifiable claim tied to JD’s actual return — not a generic “consider a K-1” recommendation a generic tax assistant would give.
The Canopy portal upload that preceded the agent build (CHANGELOG 2026-05-13) is also relevant: four W-2 PDFs (Apex Systems, GTT LLC, National Grid, Umbrage Studios) were uploaded to the tax preparer portal via Playwright automation — MFA pulled from Gmail, screenshot confirmed at ~/.playwright-mcp/canopy-w2-upload-confirmed.png. The documents that fed the agent were the same documents sent to the human preparer.
What makes document-grounded agents different
Section titled “What makes document-grounded agents different”A generic tax assistant trained on IRS publications can tell you that K-1 income exists and describe how it works. It cannot tell you whether your K-1 changes your liability by $14,300 without reading your return.
The pattern is always the same: the value of a specialist agent scales with the specificity of its grounding. A health-expert agent grounded in general fitness research is less useful than one that also knows JD’s protocol, training schedule, and bloodwork. A tax agent grounded in IRS publications is less useful than one that has parsed JD’s actual return.
This is worth making explicit because the temptation is to build the generic version first (it is faster) and add personal grounding later. The specific version is usually not harder to build — it just requires the additional step of parsing and injecting the personal documents.
How the KB is structured
Section titled “How the KB is structured”The tax strategist uses the same retrieval pattern as the professor and health-expert agents:
- A synthesis document (always loaded unconditionally)
- A goals/context file (JD’s financial objectives, loaded unconditionally)
- Top-K topic files retrieved by keyword
For tax strategy, the topic files include: the parsed return, the W-2 breakdowns, notes from the preparer evaluation, and research on applicable deductions and strategies for JD’s income composition.
The preparer evaluation is a useful addition that generic tax software doesn’t have. It answers: does the preparer we’re working with actually know about K-1 treatment, and what did they miss in prior years?
The finance agents
Section titled “The finance agents”agents/finance covers adjacent territory: financial modeling, scenario analysis, and tracking across JD’s domains (consulting income, Siemens W-2, equity/investments). The finance agents share the same specialist pattern but focus on forward-looking analysis rather than tax optimization.
The detailed CHANGELOG entries for the finance agent are sparse — it is noted in the agent registry and the package exists at agents/finance/, but it saw less development activity than the tax strategist. This is an honest gap: the tax strategist has a verified result ($14,300 swing confirmed on a real return); the finance agents are more in the research-and-scaffolding stage.
Building your own document-grounded specialist
Section titled “Building your own document-grounded specialist”Add a preparer or advisor evaluation
A document that captures what your current human advisor does and doesn’t know is high-value context for the agent. It helps the agent flag gaps.
Verify the first real query immediately
The $14,300 K-1 finding was the first test query. If the first real query doesn’t produce a concrete, verifiable claim, the grounding is too thin.
Keep money gates hard
No matter how good the recommendations get, the agent should never have the ability to execute financial transactions autonomously. Draft + approve is the right model.
Next: From structured text to voice. Plaud: Voice-Note Second-Brain Ingest covers transcription, classification, routing, and the precision problems that emerged in practice.