Open Agent Hackathon 2026 · Track 2: The Agent That Can Explain Why · 22–27 Oct 2026

Causa: an agent that explains why revenue moved, and shows its work

Brief and idea selection for Paras Adlakha. The rules require Zetaris as the data layer, a Meterless Engine in a working feature, Meterless Nightshift as the AI provider, and planner, executor and evaluator agents with critique and revision.

Recommendation

Build a revenue-variance investigator whose critic checks every number. "Why did this number move?" is asked in every monthly review. The answer always spans two systems: the ERP shows what changed, and the CRM shows why. That is a natural fit for Zetaris federation, and a question where a fluent but wrong answer is worse than none.

It matches Track 2 almost word for word: investigate a complex question across several sources and produce evidence-backed answers. Each sponsor has a load-bearing role. Zetaris joins the ERP and CRM. H-MEM remembers which dimensions explained past changes. Nightshift routes every call to Nemotron.

The problem, in numbers

35%of FP&A time spent generating insights (FP&A Trends Survey 2024)

45%of FP&A time on non-value-adding work such as data collection and validation (same survey)

2,400+finance practitioners surveyed worldwide

≈$2.6Bcloud FP&A solutions market, 2025 (Custom Market Insights)

Market-size estimates vary widely with definition: broader "financial planning software" reports range from about $4.9B to $6.8B for 2025. Treat the market figure as directional; the time-use figures come from the survey publisher's own summary.

What teams use today

Option	Examples	What goes wrong
BI "explain" features	Power BI's "Explain the increase", Tableau's Explain Data, ThoughtSpot SpotIQ	Statistical drivers inside one dataset. They don't reach the CRM for the reason, don't challenge their own findings, and don't write an answer you can cite
FP&A platforms	Anaplan, Pigment, Datarails, Vena	Strong at plan-versus-actual reporting. The narrative explaining a variance is still written by an analyst
Chat with your data	General LLM assistants over a warehouse	One-shot SQL and a fluent paragraph. No plan, no critique, no guarantee that a figure in the text matches a query
Do it yourself	Spreadsheet pivots, CRM searches, a slide	A day of work per question, and "how do you know?" has no good answer

Why this wins the track

It's the track, literally. The demo is a complex question, investigated across two sources, answered with evidence, and every step is visible.
Real agent collaboration: the Critic recomputes figures, rejects drafts and requests new SQL, so the loop changes the answer instead of decorating it.
Ground truth to check against: three business stories are planted in generated data, plus a control question with none. Judges can see the agent find causes it was never told about, and decline to invent one.
Sponsor tech is load-bearing: without Zetaris federation there is no "why", and without H-MEM the planner starts cold every time.
Safe to demo: SAMPLE_MODE and rule-based agents keep it running without services, and live runs are recorded with full traces.

Alternatives we considered

Idea	Why not
Supply-chain disruption explainer	Needs live external data (weather, shipping) that's hard to make reproducible and to register in Zetaris
Customer churn explainer	Similar shape, but real churn data is sensitive. Revenue variance is easier to generate realistically with a ground truth
General "chat with your ERP"	Fails the agent bar: one-shot SQL, no critique or revision, nothing to evaluate

Risks and how we handle them

The model invents or misquotes a number: the Critic recomputes every figure from its cited query, and the Reporter's guard rejects any figure outside the approved claims.
Correlation read as cause: confidence drops when no CRM event explains a driver, and the report says the cause isn't established (tested with a control question).
Slow hosted reasoning models: a raised first-token limit, one retry on provider errors, and a fallback to computed claims, all logged in the trace.
Laptop limits (12 GB RAM): one investigation at a time; Zetaris and Nightshift share memory carefully; a sample mode needs no services.