Countersign
The agent analyzes, a human countersigns. A personal prototype of agentic credit analysis where the agent drafts the credit memo, every figure says where it came from, and the consequential call is never the machine's to make.
Project Type
Personal
Live prototype
Timeline
Sep 2026
Team
Solo, design and build
Tools
React + TypeScript
Claude Code
What I did
- Designed and built the prototype solo, in real code.
- Made the credit memo the product. The agent drafts it, every figure carries its source, and the agent panel only reports.
- Designed the Decision card. A refusal takes two deliberate presses, and autonomy is a policy a person grants, never a default.
- Built its design system, one quiet accent under louder status colors.
Problem
Credit moves at the speed of trust, not analysis
FDIC Small Business Lending Survey, 2024. Wolters Kluwer Banking Compliance AI Trend Report, 2026.
The gap between one day and ten is the analysis. Someone has to pull the figures out of a deal document, spread them, test each covenant, rate the risk, and write the memo that defends the call. Most of that work is mechanical. None of the accountability is, because at the end a person signs.
That is why the obvious fix, letting a model write the memo, stalls in exactly this industry. An answer nobody can inspect is not an answer anyone can sign.
Insight
The signature is the product
On a clean deal, the agent's own recommendation is approve. The loop still stops. That is the name: a countersignature is the second signature that makes a document binding, and the agent's memo is only ever the first.
So the design goal was never to automate the decision. It was to make the agent's work so inspectable that a human can put their name under it without redoing it.
Checked worst first: shaky figures, then deal breakers, then anything fixable.
Why this is hard
A right answer is not enough
A provenance problem
In regulated finance, a mostly right answer with no working is a liability. Every extracted figure has to point back to its page.
An uncertainty problem
Confidence has to ride on every figure, and a critical flag has to change the agent's behavior, not just its color.
A gate problem
If approval is a dialog bolted on at the end, people click through it. The pause has to be structural.
Solution
An agent loop you can watch, ending in a signature
Countersign runs one visible loop over a deal document, and its work arrives as a credit memo you read, not a log you tail. The loop suspends itself at the consequential call.
Final Design
Nothing runs until you approve the plan
Selecting a file proposes a plan, and nothing starts until you press Start drafting. Then the memo writes itself, each section appearing when its step starts and filling in when it returns, and the run lands on the Decision card.
Every figure says where it came from
Every number carries its evidence: quoted from a paragraph of the deal document, derived from quoted figures, or not in the document at all. A click opens the source beside the memo with the figure marked.
The decision sits where you sign
The Decision card is the one place to decide. The agent's call and its sized reasons sit directly above three verbs, each with its consequence underneath. After you act, the same card becomes the sealed record.
Return it, and the agent revises
Most flawed memos are fixable, so Return for revision re-enters the loop instead of ending it. The agent searches for the smallest cure inside policy, and the cure must pass the same rule: for Atlas, $19.5M of sponsor equity, which turns escalate into approve. For Cobalt no cure fits, and the decline stands.
What would change this decision?
The stress test below the Decision card lets an analyst drag EBITDA, debt, rate, or liquidity and watch the call recompute through the same rule the agent uses. Each track is painted where the call flips, and nothing goes on the record until you attach it.
A working prototype, not a mockup
The whole product runs in the browser with no backend and no API keys, and it works on a phone, the memo stacked over the agent panel. The agent and its tools are mocked and deterministic, so every number is fictional and every run replays the same.
Design System
One quiet accent, so status stays loudest
The system has one rule: approve, escalate, and decline are the loudest colors on the page, so its aubergine accent is kept quieter than all three. Inter at 400 to 600, twelve type roles, a 4px grid, and light only.
The accent sits beside the status colors and stays quieter than all three. Every color has one job.
Inter only, twelve roles from the 22px display to the 11px meta, weights 400 to 600, uppercase only through eyebrows and tokens.
The closed status vocabulary: eighteen states, each with its glyph and its word, and the four call tokens.
One button in four variants and two sizes, every state from rest to disabled, where disabled is a token swap rather than opacity.
Reflection
Designing an agent is designing when it stops
Most agent design effort goes into making the thing act. What earned trust here was the opposite work: where it pauses, what it refuses to decide, and how completely it shows its work. The other half was layout: a reviewer should never have to ask where to look. The evidence, the Decision card, and the record are the product. The analysis is what they are wrapped around.