AWS Builder Center

Agents for Humans: keeping legal deadline math out of the LLM

Why the statutory deadline arithmetic in my eviction-defense triage agent is plain Python, what a second model found in it anyway, and where the Strands Agents SDK actually earns its place.

In Georgia, a tenant served with an eviction filing has seven days to answer. Miss it and the statute says the court shall issue a writ of possession instanter.
That word is where the project got its name.
I am building Instanter for the AWS Agents for Humans Hackathon. It watches an intake queue of Fulton County dispossessory cases and interrupts a supervising attorney for the handful closest to a default judgment. Eviction-defense clinics are walk-in operations, and somewhere in that stack of paper are the two or three people whose deadline is tomorrow.
The first architectural call is the one I want to write about. The deadline math never touches the model.

Why I did not let the model near it

AWS Prescriptive Guidance puts it plainly: "Use deterministic execution logic unless AI is needed. Implement coded logic for deterministic execution. Use AI functionality where it provides value that deterministic code cannot."
A statutory deadline is the textbook case. O.C.G.A. 44-7-51(b) gives seven days from actual service. O.C.G.A. 1-3-1(d)(3) makes the day of service day zero and rolls the terminal day past Saturdays, Sundays and legal holidays. There is no judgment call anywhere in that.
An LLM doing this arithmetic would be slower, unauditable, and wrong some fraction of the time. In this domain that fraction is a family losing a home.
So the engine is plain Python. A jurisdiction rule table, two calendar sets, a day counter, and a full per-case trace. The Strands agent wraps around it and does the two things a model is genuinely good at here: reading messy intake notes into typed fields, and explaining in plain language why one case outranked the others for an attorney's scarce attention.

The part I did not see coming

The deterministic half turned out to be the hard one.
Georgia's legal holidays are the federal holidays as of January 1, 2022, plus whatever the Governor proclaims that year. The Fulton County courthouse publishes its own closure calendar. You would assume those match. For 2026 they disagree on exactly three dates.
April 3 is a state legal holiday and the courthouse is open. October 12, same thing. Those two mean the statute rolls a deadline off a day when the clerk would happily accept a filing.
December 31 is the dangerous one. The courthouse is closed and it is not a legal holiday. A tenant served on December 24 has a day-seven deadline of December 31. The statute does not move it. The building is locked.
My engine will not resolve that. It computes the date, raises a flag saying a human has to decide, and tells you when the clerk's office reopens only if the encoded calendar actually covers that date. Past its coverage it returns nothing rather than a guess.
Every ambiguity becomes a flag for a person. Never a quiet resolution.

Then a second model went through my code and found the fail-open paths anyway

I run an adversarial review before anything ships. Round one found three real defects in a suite that was already 37 tests green.
When calendar coverage ran out and the computation refused, a summons-stated date could still flow into the result looking like a finished deadline. Every deadline now carries a provenance enum saying what it rests on, and an unverified summons date is labeled exactly that.
The "when does the court reopen" advisory would guess that dates past its closure data were open days. It returns nothing now, the moment the data runs out.
The third one is the one that bothers me. A malformed rule row, the string "next_non_weekend_non_holiday" sitting where an enum belonged, silently skipped the roll and produced a Saturday deadline with no warning at all. Rules validate at construction now and the dispatch raises on anything it does not recognize.
Round two then found defects in my fixes, including the old Python trap that isinstance(True, int) is true, so a boolean could quietly become a one-day window.
The loop does not stop when I declare it finished. It stops on a clean round.

Where Strands earns its place

None of that needed an agent, which is sort of the point.
The agent layer is built on the Strands Agents SDK and handles what plain code cannot: a Graph that sequences perception, computation, ranking and escalation, a human-in-the-loop interrupt that pauses the run and hands one case to a licensed attorney with a rationale attached, and session persistence so that decision can arrive days later.
The model perceives and communicates. The Python decides dates. The attorney decides everything that matters.
Next post: how the escalation ladder rations attorney attention without collapsing into a score, and why the interrupt is the product.
Instanter drafts only and never files, computes deadlines and never advises, and keeps a licensed attorney in the loop for every case it surfaces. All demo data is synthetic. The statutes are real and public. Built solo for the Agents for Humans Hackathon.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article