Question 1

What actually is an AI agent, in plain language?

Accepted Answer

An AI agent is an AI model that works toward a goal by deciding its own steps and using tools, where a chatbot answers one message and waits for the next. Anthropic draws the line this way: in a workflow, code fixes the steps; in an agent, the model directs its own process and tool use. In practice an agent opens a browser, reads your email or files, runs code or edits a spreadsheet, checks the result and carries on until the job is done or it needs you, as ChatGPT Work, Claude and Tasks in Gemini now do. For you it means handing over a whole task rather than a question, then checking what it did before trusting it with money or messages.

Question 2

Why would a founder use agents instead of a chatbot or automation?

Accepted Answer

Chatbots answer questions and classic automation follows fixed if-this-then-that rules; both break the moment a task needs judgment. Agents sit in between: they read context, decide, and take multi-step action across your tools, so one agent can replace a pile of brittle zaps plus the human glue between them. Use automation for fully predictable flows, a chatbot for Q&A, and an agent when the work requires decisions along the way.

Question 3

What real tasks are founders delegating to agents today?

Accepted Answer

Founders are handing agents inbound sales and lead qualification, customer support, competitor and market research, meeting prep and follow-ups, content drafts, ops reporting, and a growing share of coding. SaaStr's AI inbound agent closed over $1M in 90 days, and 15-person Every runs multiple products with agents doing most of the execution. Agents now also hold standing jobs, like watching a pipeline or preparing the Monday report. The pattern: repetitive, text-heavy work with a clear goal goes first, and the founder stays on as reviewer.

Question 4

How do I build my first agent without being technical?

Accepted Answer

Start inside the assistant you already pay for, and reach for a builder only when the job has to run across several apps. On Claude Pro ($20 a month) you can hand off a task in plain English and schedule it to repeat, and Tasks in Gemini does the same on Google AI Pro, ₹1,950 a month in India. To join your apps, Zapier (free for 100 tasks a month) has Agents, now moving into AI by Zapier; Lindy works from Slack and email with $50 of free credits for 7 days, then from $29.99 a month; and n8n suits long workflows or running it on your own server. If you do write code, OpenAI's Agents SDK, Anthropic's Claude Agent SDK and Google's ADK are the standard libraries. Give the first agent one narrow, low-risk job, such as labelling your inbox or drafting meeting briefs, watch every run for a week, then widen what it can touch.

Question 5

Which platform or framework should I build my agents on?

Accepted Answer

Match the tool to who is building. Non-technical founders should start no-code (Zapier Agents, Lindy, n8n) and only graduate when they hit a wall. If you have engineers, OpenAI's Agents API (also offered as Amazon Bedrock Managed Agents) now runs the loop, the sandbox and computer use for you. To own the loop, LangGraph gives the most control, CrewAI is fastest for role-based prototypes, and the OpenAI Agents SDK is the lightest. Whatever you pick, start with the simplest thing that works; most agent products are one model, one loop and good tools.

Question 6

How much do AI agents cost to run, and how do I keep the bill under control?

Accepted Answer

Token prices for equivalent quality fall roughly 10x a year, yet agent bills keep rising because agents burn about 50 times the tokens of a chat: every step re-sends the whole history. A simple task can cost a fraction of a cent while a long agent session runs several dollars. The levers that matter: prompt caching (up to 90% off repeated context), routing easy steps to cheap models (GPT-6 Luna, Claude Haiku 5.5), sending yes/no steps to a decision model like Jev, trimming history, and measuring cost per completed task rather than per token.

Question 7

How do I test an AI agent and know when it's reliable enough to trust?

Accepted Answer

You test agents with evals: a set of real tasks plus grading logic, run repeatedly because the same input can produce different outputs. Start embarrassingly simple: read 50 transcripts of your agent working, label every failure, turn the recurring ones into automated checks (code checks where possible, an LLM judge validated against your own labels where judgment is needed). Trust is earned per task: track pass rates over many runs, not one good demo.

Question 8

What can go wrong when an agent acts on my behalf, and how do I keep it safe?

Accepted Answer

The core risk is Simon Willison's lethal trifecta: an agent that reads your private data, sees untrusted content (any email or web page) and can communicate externally can be tricked by a poisoned message into leaking or destroying things, with no bug in your code. Prompt injection remains unsolved, so design around it: minimum access, never all three capabilities in one agent, human approval for anything that sends, spends or deletes, and full logs. Claude in Chrome and ChatGPT's dots now act without asking, a second model checking each step; treat that as a backstop, not your approval.

Question 9

What is MCP, and why does it matter for my agents?

Accepted Answer

MCP (Model Context Protocol) is an open-source standard for connecting AI applications to outside tools and data, which its own documentation compares to a USB-C port for AI. An MCP server is a small program that exposes one system, such as GitHub, Notion, Gmail or your database, as tools and data; the AI app that uses it (Claude, ChatGPT, Cursor, VS Code) is the client. Anthropic released it in November 2024 and the current specification is dated 28 July 2026. For you it means one connector works in every assistant that speaks MCP: Claude lets even Free users add one custom connector from a URL. Each server widens what the AI can read and do, so connect only servers you trust.

Question 10

Should I build one agent or a team of agents working together?

Accepted Answer

Start with one. The famous industry debate (Cognition's "Don't Build Multi-Agents" vs Anthropic's multi-agent research system) actually agrees on the fundamentals: multiple agents working from partial context make conflicting decisions, and multi-agent setups burn roughly 15x the tokens. Cognition's 2026 follow-up narrows the rule: extra agents can research or review, as long as only one writes. Go multi-agent only for parallel, read-heavy work like broad research, with one lead agent delegating to sub-agents.

Question 11

Can coding agents like Claude Code actually build my product?

Accepted Answer

Yes, further than most founders expect: teams like Every ship multiple products with engineers writing almost zero code by hand, and 80-90% of Claude Code itself is written by Claude. Coding agents now also run in the cloud, so you can hand one a ticket and close the laptop. The catch is the last stretch: agents get you a working MVP fast, but edge cases, security and production hardening still need engineering judgment, and AI-generated code often ships vulnerabilities if nobody reviews it. Treat the agent as a very fast builder whose work you spec, review and test: Lovable or Bolt if you never want to see code, Claude Code, Codex or Cursor if you or a teammate will.

Question 12

Is there a real startup opportunity in building vertical AI agents?

Accepted Answer

Yes, and it may be the biggest one of this cycle: YC's partners argue vertical AI agents could be 10x bigger than SaaS because they tap labor budgets, not software budgets; you sell completed work, not seats. The winners (Harvey in legal, Sierra and Decagon in support, EvenUp in personal injury, Abridge in healthcare) pick one painful workflow in one industry and own the outcome end to end. The edge goes to founders who deeply know a niche's workflow, including unglamorous Indian and SMB verticals nobody in San Francisco understands.

Question 13

How should I price a product built on AI agents?

Accepted Answer

Per-seat pricing breaks when your product replaces labor instead of assisting it: agents don't have seats. The market is converging on usage and outcome pricing (Fin, formerly Intercom and now part of Salesforce, charges $0.99 per outcome; Zendesk bills per automated resolution; Salesforce charges $2 per conversation), usually wrapped in a hybrid: a platform fee plus variable consumption. Anchor your price to the value of completed work, agree contractually on what counts as an outcome (Fin bills a qualified lead at $9.99), and watch your token costs so an unlucky customer doesn't make you lose money per task.

Question 14

Why do most AI agent projects fail, and what do the winners do differently?

Accepted Answer

MIT's much-cited finding is that 95% of enterprise AI pilots deliver no measurable ROI, and the cause is rarely the model: it's missing context, no workflow integration, and celebrating the demo instead of the deployment. LangChain's industry survey adds that quality is the top blocker and teams with observability and evals ship far more reliably. The 5% that win start with one narrow, high-value use case, wire the agent into real systems and data, measure outcomes from day one, and keep a human on the risky steps.

Question 15

How do I give my agent the right context and memory so it stops forgetting?

Accepted Answer

Agents fail less because the model is dumb and more because its context window is stuffed with the wrong things: performance measurably degrades as context grows (context rot). Context engineering is the fix: write important facts to external notes, select only what the current step needs, compress long histories into summaries, and isolate big sub-tasks into fresh contexts. Manus's production lesson adds a money angle: structure context so it caches well, because cache hits are 10x cheaper than fresh tokens.

How founders use AI for AI Agents

The questions

Also in Starting Up