Seven working agent builds with real stacks and real code paths: document Q&A, email triage, research briefs, support routing, code review, meeting actions and a news digest.

AI Agent Examples 2026: 7 Real Working Agents You Can Build This Weekend

An agent is a loop with tools: the model decides, calls a tool, observes the result, repeats until done. Frameworks change, the loop does not. The fastest way to learn is to build one real thing.

1. Document Q&A agent. LlamaIndex (or LangChain loaders), a real embedding model, local or API LLM. Load, chunk, embed, retrieve top-k, answer only from retrieved context. Roughly a hundred lines. Do not add multi-query and reranking before the basic path works.

2. Email triage agent. Gmail API or IMAP plus a small model with a fixed label set, returning JSON of label, urgency and one-line summary. Heavy mail users save 30-60 minutes a day. Keep labels to a closed enum, or you will get twenty invented ones.

3. Research agent. LangGraph or LangChain, a search API, a scraper and a long-context model. Split the question, search per sub-question, clean the page text, dedupe, then write a brief with source links. Scraping quality matters more than model choice.

4. Support routing bot. n8n or Dify with a knowledge base and a ticket API. Classify intent, answer when the knowledge base hits, otherwise open a ticket with the conversation summary attached. Always define an explicit handoff condition.

5. Code review assistant. Aider or Cline driven from GitHub Actions. Feed the diff plus the call sites of changed functions, not the diff alone, and filter out style nitpicks.

6. Meeting action extraction. Whisper or WhisperX locally, then an LLM that outputs tasks as JSON. Only accept owners explicitly named in the transcript, otherwise the model invents them.

7. Personal news digest. RSS plus an LLM plus a scheduler. Dedupe by title similarity and URL before summarising, or the same story appears eight times.

Three recurring mistakes: over-engineering (five tools failing silently beats one working tool), no human checkpoint before irreversible actions like sending or ordering, and unbounded loops without iteration caps, timeouts or a per-task budget. Start with the research agent: highest value over complexity, no internal systems to integrate.