# Agentware > Agentware is a free, online, 72-hour hackathon by Hackathon Raptors (November 27-30, 2026) for building AI agents that do real work on their own and can be trusted to run unattended. Software only. Theme: build an agent you can walk away from, then prove it knows when to stop. - Website: https://agentwarehack.com/ - Register: join the Hackathon Raptors Discord, https://discord.gg/xfYPDZYqeh - Organiser: Hackathon Raptors, https://raptors.dev, hello@raptors.dev ## Key facts - Dates: November 27, 2026, 18:00 UTC to November 30, 2026, 18:00 UTC (72 hours) - Format: online, free, software only (no hardware) - Teams: 1 to 4 people, solo welcome - Prize pool: $1,800 ## Tracks - Inbox & Calendar: An agent that runs your email and your schedule: inbox triage, replies and follow-ups, scheduling, a daily digest. - Life Admin: An agent for the chores that pile up at home: bills and renewals, forms and paperwork, travel planning, groceries. - Research & Study: An agent that watches the sources you care about and keeps your reference library, notes and deadlines in order. - Browser Agent: An agent that works the web the way a person does: navigating, filling in forms and extracting data on sites with no API. - Developer Workflow: An agent for the chores between writing code and shipping it: PR triage, code review, failed-build diagnosis, on-call first response. - Open Track: Build your own. No limit on domain, input or stack, as long as it runs unattended. ## Prizes - 1st place: $800 - 2nd place: $400 - 3rd place: $200 - Best Kill Switch: $100 - The Near Miss write-up: $300 (three awards of $100) ## Scoring Each project is rated on a 5-point scale across five weighted criteria. A safety bonus adds up to +4: guardrails, live monitoring, kill switch, audit trail (+1 each). - Autonomy (30%): How much of the job runs with no human in the loop, verified by the trace and the demo, not the README. How many steps run without a human. Whether it plans, checks and finishes on its own. - Usefulness (25%): Does it do a job someone would genuinely hand over? Real problem, real workflow, real time saved. A narrow agent that nails one job beats a general one that half-does ten. - Reliability (20%): What happens on bad input, a failed API call, a timeout, an ambiguous instruction, a trap pack email? Does it recover, retry, ask, or stop cleanly? Or does it carry on confidently in the wrong direction? - Engineering (15%): Clean architecture a senior reviewer would respect. Sensible tool design. Permissions scoped to what the job needs. Docs a stranger can follow without asking you a question. - Demo (10%): Does the video show one real, unattended run from trigger to result, clearly enough that a judge believes it before running it? ## What a submission needs - A judge can run it: clone, follow the README, fill in .env.example, working in under 15 minutes. - A judge can watch it: a demo video (3 minutes max) of one full unattended run, trigger to result. - A judge can read it: every run leaves a trace of what it saw, decided, called and got back. - Trap Pack: hostile test inputs (prompt injections, conflicting instructions, broken data) released at kickoff; commit your results. ## Timeline - October 26, 2026: Registration opens - November 6, 2026: Judging panel announced - November 23, 2026: Team formation closes - November 26, 2026: Full brief and scoring rubric published - November 27, 2026, 18:00 UTC: Kickoff, Trap Pack released, hacking begins - November 30, 2026, 18:00 UTC: Code freeze, submissions due - November 30 to December 10, 2026: Judging window - December 7, 2026, 18:00 UTC: The Near Miss write-ups close - December 11, 2026: Winners announced ## Rules - Open source under an OSI-approved license (MIT or Apache-2.0 preferred); you keep full ownership. - Any model, any stack; name what you used in AGENT.md. - The agent must take real actions through real tools, against real or sandboxed systems. - A stranger can set it up from the README in under 15 minutes. - No secrets in the repo; committed secrets disqualify. - Do no harm: no spam, no unsolicited messages to real people, no terms-of-service violations, no real money without hard limits. - All project code written during the 72-hour window. - Teams of 1-4 people; solo welcome. - AI coding tools are expected. ## FAQ ### What's the team size limit? 1-4 people. Solo welcome. ### Is there a theme or track I have to pick? Pick one of six tracks. If nothing fits, the Open Track has no limit on domain, input or stack. Every track is scored on the same rubric. ### Can I use hosted models like Claude, GPT or Gemini? Yes. Local models too. Name what you used, and note roughly what a judge will need to run it. ### Do judges need to pay for API keys to test my agent? Keep it cheap. Free tiers, a mock mode or a recorded trace for expensive paths are all fine. Say clearly in your README what a test run costs. ### What if my agent needs access to my email, calendar or accounts? Build it against a test account or a sandbox, and document how a judge can do the same. Never ask judges for their own credentials. ### Does voice get its own track? No. Voice is one input among many, so a voice agent enters the track its job fits, or the Open Track. Every track is scored on the same rubric. ### What is the trap pack? A set of hostile test inputs released at kickoff: prompt injections, conflicting instructions and broken data. It feeds into Reliability. You do not need to pass all of it, but you should know what your agent does when it fails. ### Is the safety bonus mandatory? No. It is up to +4 points on top of your score. Given how close final rankings usually are, we would not skip it. ### Can I build on an existing agent framework? Yes. LangGraph, CrewAI, the Agent SDKs, MCP, n8n, all fine. A pre-existing project of yours, or an existing open-source agent with the name changed, is not. ### Can I start coding before November 27? No project code. Planning, sketching prompts, picking your stack, setting up test accounts, all fine and encouraged. ### Do I keep ownership? Yes, entirely. Nothing to sign. ### Why "an agent you can walk away from"? Because everyone wants the assistant that quietly handles everything, and nobody wants the one that decides it knows better. The difference is engineering. That is what we are judging. ## Judging at Hackathon Raptors Every Hackathon Raptors event is judged by a panel formed for that event and matched to its subject. The panel for Agentware is being formed now. To nominate someone, or to be nominated, write to hello@raptors.dev with a short note on your background. - How judges are selected: https://www.raptors.dev/judges - Contact: hello@raptors.dev ## Upcoming Hackathon Raptors events - Code Olympics Winter Championship 2026 (18-21 December 2026): https://codeolympics.com/winter-2026 - Every event, past and upcoming: https://www.raptors.dev/llms.txt