Dwelly London AI Hackathon · 10 October 2026

What AI agents can already do in a letting agency

Twenty-seven teams had one day to build agents that run the messy middle of a lettings business - repair requests, landlord approvals, contractor chasing, invoices - and then face cases they had never seen. These are their submissions, with the demo videos the teams recorded on the day.

The hackathon

Hosted by Dwelly, the AI-native UK lettings group, with Anthropic, ElevenLabs, General Catalyst and EQT as partners. Places were selected from applicants: 424 registered, about 75 made it into the room, many with proper product backgrounds.

424registered
~75made it into the room
27teams submitted
4.5hof build time

The test

Each team got a synthetic operations environment - calls, emails, documents, CRM records, tickets - and picked a track: Property Management, Insurance Claims, Banking, or Delivery and Logistics. Thirty minutes before the deadline, an unseen "Reality Test" arrived: cases with conflicting information and missing context, designed to see whether an agent knows when to stop and hand over.

It wasn't a blue-sky ideas contest. Every team worked the same specific cases, and each agent was run against cases it had never seen and scored on whether it got each one right - which is what makes the results telling.

How it was judged

  • 35% correct outcomes
  • 20% coverage of completed cases
  • 20% judgement and safety handling
  • 15% engineering quality
  • 10% real-world applicability

More than half the marks were for being right and knowing when to escalate - not for speed.

Six minutes: the best of the day

The four prize winners and our own entry, back to back.

The lettings submissions

Property Management track. Each video is the team's own one-minute demo.

Grand prize · Best Real-World Agent

Monopoly - Casework

One repair inbox for the agency - phone, email and web form - with an AI agent proposing the next step on every case and a person on every decision that matters.

The clever bit: simple tenant replies go out on their own; anything to a landlord or contractor, anything involving money or a safety risk (gas, CO, electrics, flooding) waits for a person. In the Reality Test it held an invoice and flagged an unverified change of bank details.

github.com/Harrishayy/monopoly

Best Engineering

Properly

"Every call, email, invoice and bank line about a property, in one graph an agent can act on."

The clever bit: one record per person, property and contractor across every source - "one 'Moss Plumbing & Heating', however it is spelled" - with every fact linked to the document it came from, and books that are added to, never edited. The foundation for merging agencies onto one system.

github.com/Fergus-MW/propadm.in

Our entry · Freddie Chambers and Isabella Isotta

RepairDraft - No repair goes quiet

Reads every open repair case, decides the next step - act, ask, wait or escalate - and drafts the messages to tenant, landlord and contractor.

The clever bit: a written rulebook checks every decision in code and can only ever make it more cautious, never less. Around it: a coordinator dashboard, live email and text intake, and a landlord bot for approvals. All 30 Reality Test cases answered, no safety failures.

Property Management

D&D - Lettings Ops Agent

A maintenance desk run from a rulebook, where every exception reaches a person as a 90-second decision card: the question, the gap, the options, a recommendation, and what happens if nobody answers.

The clever bit: they found that a default spending limit let the agent spend money even after it had spotted a fake approval - so they removed standing authority altogether.

github.com/dkushnikov/lettings-ops-agent

Property Management

Fabriq Mafia - Triage and Next Action

A triage engine for anything that comes in - email, call or ticket - that decides what happens next and why.

The clever bit: the model proposes, code checks before anything runs, and the record is written from what actually happened. Urgent cases page the on-call manager by WhatsApp and phone, and the reply feeds straight back in.

github.com/jeremy-shr/dwelly-hackathon

Property Management

Agents of Estate - Maintenance Case Worker

An autonomous case worker with an ops portal, a one-page brief for every escalation answered over WhatsApp, and a voice agent that phones landlords for approval.

The clever bit: every action is read back and checked before it counts; bank details never change because a message says so, and "ok" or an emoji is not an approval. About 8p and 28 seconds per case.

github.com/martin-dutchy/dwelly-agent

Property Management

CaseAgent - Northbridge Lettings

Maintenance tickets for a fictional agency, with saved state so a case picks up again when the landlord or contractor replies.

The clever bit: a small model reads each landlord reply as approved, refused or unclear, and the agent refuses to send any message quoting an amount or date that isn't in the case file.

github.com/hridaya423/caseagent

Winners from the other tracks

Built for delivery desks, but the techniques transfer straight to tenant calls and contractor chasing.

Best Voice Agent · Best Exception Handling

Gingerbread King - Great Agent

Answers delivery-desk calls and decides the next safe step.

The clever bit: the caller only ever hears pre-approved lines, so it can never promise money or dates, and any hazard word triggers "call 999 now" first. An independent reviewer checks every plan before anything is sent.

github.com/workonthestreets/greatagent

People's Choice

Silly Hamsters - Hamyway City

Watch the agent think: every case plays out live across a simulated city, with playbooks you can see and edit.

The clever bit: making an agent's decisions visible and legible to the people supervising it - the room voted it their favourite.

github.com/arvin-shafiei/hamyway-city

What every lettings team converged on

Working independently, the teams arrived at the same design - which is probably the clearest picture yet of how an AI-run letting agency front office works.

  1. An agent on every case, one thread across tenant, landlord, agent and contractor. Nothing goes quiet because nobody owned the next step.
  2. Hard rules live in code, not in the prompt. The model proposes; deterministic checks decide what is allowed to happen.
  3. Landlord approval is the control point. No contractor and no spend without a verifiable yes - and a thumbs-up emoji doesn't count.
  4. Escalation is a product, not a failure. When a person is needed, they get a ready-made brief: what's known, what's missing, the options, a recommendation.
  5. Everything leaves a record. A reason and a trace for every decision, so the work can be audited and improved.

Every team built for maintenance, and most held back or escalated a large share of the unseen cases. The near-term prize is more homes per property manager, not a fully autonomous agency. But it is about the most agent-ready use case there is, in a very fragmented market - and if it isn't showing up in revenue yet, it will be soon.

Why Dwelly ran it

Dwelly buys independent UK letting agencies and runs them on its own AI platform. Founded in 2023 by former Uber and Gett operators, it has raised over $260m including debt - most recently a $170m Series B in July 2026 ($95m equity led by EQT Growth, plus a debt facility) - and now runs around 20 agencies and 16,500 homes.

Every submission had to be published as open source under Apache 2.0, so for a modest prize pot Dwelly can learn from 27 approaches to its own operational problems, and use the cases to benchmark its own agents.

Dwelly's published numbersBeforeAfter
Homes per property manager~100300+
Human time per maintenance job~3 hours15-20 minutes
Time to let~3 weeksunder 2 weeks
Payments matched automatically-80%+

Sources: Dwelly Series B · Dwelly AI operations · Fortune, Feb 2026 · Dwelly agencies