AI Agent Examples: 40 Real Agents Built in 2026

Most lists of AI agent examples show the same famous products. This one shows 40 agents that individual builders and small teams made in 2026 and submitted to the OpenAI hackathon on Devpost. We went through the more than 1,000 submissions that describe themselves as agents, picked 40 that show the range (most of them liked by other participants), and grouped them by the job they do: support and sales, back office, coding, personal life and specific industries. You can browse all of them in our archive of 7,800+ hackathon projects.

One caveat before the list: these are hackathon projects. The descriptions are the builders' own, taken from their Devpost taglines and write-ups, and nobody has independently checked that each agent works as described or is used in production. Read them as examples of what people are building and as ideas worth testing, not as product reviews. Each name links to the idea page with a longer summary.

Updated October 2026. All counts in this article come from our copy of the Devpost gallery, captured on 13 August 2026.

What counts as an AI agent (and what doesn't)

An AI agent is software that takes a goal, decides which steps to take, uses tools to take them (a browser, an API, a database, a code sandbox) and checks the result before it stops. Anthropic's engineering guide Building effective agents (19 December 2024) draws a useful line: workflows are systems where language models and tools follow predefined code paths, while agents are systems where the model directs its own process and tool use.

• Chatbot: answers questions in a conversation. It becomes an agent when it can act, for example book the appointment instead of explaining how to book it.
• Workflow automation: runs the same fixed steps every time (a form arrives, a CRM record is created). Plenty of products sold as agents are well-designed workflows, and that is often the right choice.
• Agent: picks its next step based on what it finds, can loop and retry, and stops when the goal is met or when a person needs to decide.

Most projects below sit between the last two. A model plans and acts, ordinary code does the arithmetic and enforces the rules, and a person approves anything risky, such as payments, emails to customers or code merges. That human-in-the-loop pattern comes up again and again in the list.

Customer support and sales agents

Front-office agents answer, qualify and follow up: jobs where speed matters and the same questions keep coming.

• ChatNEX: turns a WordPress site into a sales and support assistant that answers from the business's own content, collects leads and keeps up with changes such as prices and stock. Its builder says they are not a programmer and made it with ChatGPT.
• ListingOS: a seller agent for eBay. You photograph products, GPT-5.6 drafts price-aware listings, and you publish them from one review screen.
• CallChat ZERO: a lead-capture chat agent built around consent. It collects only what the visitor agrees to share, limits how long data is kept and hands over to a person with an audit trail.
• TeleVoice: a Telegram bot that places outbound AI sales calls in English or Spanish through Twilio. You message the bot, and it calls the number and runs the conversation.
• Agentic-CS: a multilingual customer-service platform for small businesses, built with Malaysian SMEs in mind, that answers from approved business knowledge, follows controlled workflows and keeps a person in charge.
• Briefly: turns a prospect's company name into a discovery-call brief with company context, likely pain points, tailored questions and a timed agenda.
• ChurnControl: a retention dashboard for customer-success teams that flags accounts at risk of leaving, explains why and proposes actions a person approves.

The pattern worth copying: each one limits the agent to one kind of conversation and makes the handover to a human explicit. For more ideas in this area, see AI ideas for marketing and sales.

Operations and back-office agents

Back-office agents handle reconciliation, triage and paperwork. The work is slow and full of rules, and mistakes cost money, so these builders lean hardest on verification.

• flowpilot-ai: seven coordinated agents that triage an inbox, manage tasks, schedule meetings, handle support tickets and write an executive report from a single instruction.
• Meeting-to-Action Agent: turns a meeting transcript or audio recording into action items with owners, then creates the matching GitHub issues.
• SupplyTact: a supply-chain planning agent, designed around pharmaceutical operations, that investigates supply risks, runs deterministic simulations and shows the evidence behind each recommendation.
• Argus: a crew of agents that runs vendor due diligence, including on AI tools, agents and MCP servers, for companies without a security analyst.
• AuditGuard DE: reconciles German accounting reports (EXTF, SuSa, BWA and USt-VA) and stops with a list of missing evidence instead of guessing values.
• Ledger Sense AI: matches messy bank payments to invoices. GPT-5.6 resolves ambiguous names and references, ordinary code does the matching maths and posting, and every decision is explained.
• ARIA: a revenue-intelligence agent for small Nigerian merchants that reads WhatsApp messages and bank alerts overnight and delivers three prioritised actions each morning.
• Match2: a recruitment agent that matches candidates to jobs on skills, seniority, industry and must-have requirements, ranks them and explains each match.

Notice how often the model is kept away from the numbers. AuditGuard DE and Ledger Sense AI both let deterministic code do the arithmetic and use the model only for the fuzzy parts, a sensible split for anything that touches money. More finance projects are collected in AI ideas for finance.

Developer and coding agents

Software is where agents are most common in this corpus: by our keyword count, 364 of the 1,032 agent projects mention code, developers, repositories, bugs or CI in their name or tagline. Some are agents that fix code; many are tools that watch, record or limit other agents.

• Hermes World: a real-time visual observability layer that shows what each agent in a multi-agent system is doing. With 22 likes, it was the most-liked agent project in the gallery.
• RuntimE.Razor: helps coding agents remove dead code using runtime evidence, meaning which code actually ran, instead of static analysis alone.
• Autonomous CI Failure Repair Agent: reproduces a failing CI run, writes a fix and a regression test, verifies both in a clean sandbox and opens a pull request only when every check passes.
• Cortex: turns a solo agent chat into a team session. Teammates mark up agent-built UI on a shared canvas, and the agent applies the changes to the code.
• AutoFix Swarm: a multi-agent system that diagnoses software failures, generates targeted fixes and verifies each patch before handing it over.
• GlassHouse: a flight recorder for coding agents. It records file edits, terminal commands and test runs, then replays the session with tamper-evident hashes.
• ReproBot: takes a GitHub issue, writes a minimal reproduction script, runs it and shows whether the bug is real.
• War Room: a team of incident-response agents that diagnoses an outage, finds the commit behind it, proposes a fix and drafts the postmortem, with each agent's reasoning visible as it works.

The number of guardrail tools is a signal in itself: people who use coding agents every day want records, replays and approval gates. Whatever domain your agent is for, plan that layer from the start.

Personal and life agents

Consumer agents act on someone's behalf in daily life: at home, on a computer, in the kitchen or on holiday.

• Sara Smartest Speaker: a smart speaker built on Raspberry Pi hardware that holds a two-way voice conversation and hands tasks to an open-source agent with tools and skills. With 14 likes, it was the second most-liked agent project.
• Porter: turns a Mac into a voice-controlled agent that sees the screen, uses your apps and learns your preferences.
• Clark: a web agent that completes browser tasks from plain language, including searches, forms and logins, and leaves anything sensitive to you.
• AJÒ Beach: plans a beach day in Sardinia and re-plans it as wind, waves and your own constraints change.
• Tail Villain: turns a personal goal into a roadmap, with progress analytics and live voice coaching.
• BepFlowAI: helps you decide whether to cook, meal-prep or eat out, with separate agents for recipes, restaurants and what is already in your kitchen.
• Nook: remembers your home. Photograph the fridge or tag an item, and it tracks expiry dates and locations so it can answer “is my milk still good?” or “where's my passport?”
• GenkanAI: turns an existing home intercom into an agent that talks to visitors, using an add-on ESP32 board instead of new wiring.

Personal agents depend on trust and on how little setup they need. The clearest entries here start with one routine and one device rather than trying to run your whole life.

Industry agents: health, finance, legal and property

Industry agents carry rules that general-purpose tools don't know. In regulated fields, these builders consistently keep a professional in the approval seat.

• VitalCue: guides someone through first aid in real time by watching and listening through the phone. It is described as an offline-first Android prototype.
• AIHealthExpert: a clinical workspace where GPT-5.6 agents handle analysis, evidence synthesis and documentation, with clinicians in control of decisions.
• Autonomous Wellness Front-Desk Agent: a 24/7 chat agent for clinics that answers common questions, checks live slots, books and reschedules visits and passes the conversation to staff with full context.
• CREDGEN-AI: a multi-agent lending platform for non-banking financial companies (NBFCs) that covers customer engagement, underwriting, fraud checks and sanction letters.
• Served: helps small businesses answer financial subpoenas without a legal team. It checks whether the request is genuine, pulls the matching records and drafts the evidence package for human review.
• LexStrike: stress-tests a contract by having hostile agents attack each clause while a defender agent argues back, then lets you fix and re-check it.
• MaLu Estate AI Agency OS: a system for a real-estate agency in which agents prepare lead follow-ups, listing copy and internal briefs while people handle the moments that need trust. Its founder says it is being developed alongside a working agency in Croatia.
• MoveOutDocs: helps small landlords document damage at tenant move-out with structured photos and wear-and-tear rules for their jurisdiction, then produces an itemised deduction letter.
• Umbra: plans daily sun-safety schedules for outdoor work crews from live UV, weather and site-photo data across several job sites.

Health ideas have their own collection: AI ideas for healthcare.

How these agents were built

Devpost asks every team to list what they built with. We counted those tags across the 1,032 projects whose name or tagline mentions an agent, agents or agentic; 1,029 of them list at least one tool. The tags are self-reported, and because OpenAI ran the hackathon, its own tools are over-represented.

Three things stand out. Most agents run on ordinary web stacks (TypeScript or Python, React, a SQL database), not exotic infrastructure. Agent frameworks are rare: LangGraph appears in 33 projects, LangChain in 18 and the OpenAI Agents SDK in about 15, so most builders wrote their own loop around a model API. And MCP, the open protocol for connecting models to tools and data, is already listed by roughly one agent project in seven.

How to build your own AI agent MVP

Most of the agents above were built by one or two people during a hackathon. Turning one into something people pay for is mostly a matter of narrowing it down. A practical order of work:

1. Pick one job with a clear finish line. “Match last month's bank payments to invoices” is an agent; “run my finance team” is not. Write down what done looks like and who checks it.
2. Start as a workflow, then add autonomy. Script the steps you already know and let the model decide only where the path really varies. Anthropic's guide gives the same advice: find the simplest solution possible and add complexity only when it is needed.
3. Give it a few well-defined tools. Each tool (search the CRM, read a calendar, run the tests) needs clear inputs, clear error messages and the narrowest permissions that work. MCP is a common way to expose them.
4. Keep maths and rules in code. Calculations, validation and anything legal or financial belong in deterministic code, as AuditGuard DE and Ledger Sense AI do it.
5. Put a human gate on irreversible actions. Payments, outgoing emails, merges and deletions wait for approval until the agent has earned trust.
6. Log every step and test against real cases. Keep a record of each tool call and decision, the GlassHouse idea, and build a small set of real tasks with known right answers to rerun after every prompt or model change.
7. Watch the cost per task. An agent that loops calls the model many times. Measure tokens per completed task early, before pricing the product.

For the wider build process for any AI product (choosing a model, retrieval, evaluations and guardrails), see how to build an AI app. Running costs such as tokens and monitoring are covered in AI app development cost.

Before writing code, check that someone wants the agent at all: how to validate a startup idea walks through cheap tests. A clickable prototype of the approval screens is often the quickest way to show users what the agent will do and where they stay in control; AI prototyping tools lists the options.

Frequently asked questions

Have an agent idea that needs a first version? We build custom MVPs on monthly plans: Bi-weekly sprints at $399/month, Weekly sprints at $899/month, or 24/7 development, priced on request. See the pricing page for what each plan includes.