Every agent on this page actually ran. Here's what each one cost — and how each one broke.

What counts as a real AI agent example?

The bar on this page: a run date, a cost, and a failure mode. Without all three it's a demo — and the internet does not need another demo. (Definitions live at what is agentic coding; this page is what the definitions look like with money attached.)

What does a multi-agent swarm look like in practice?

1. The book swarm

15 agents, 25,000 words, ~6 minutes, under $40. One dispatch message, each agent writing one chapter to a file, a verification pass after. The artifact you're reading started this way (Ch 6).

2. The strategic-plan swarm

Specialist agents fan out against a locked brief, then a reconciler and a red-team attack the result. Two runs on record: 20 documents in 11 minutes for $6.20, and 25 documents in ~14 minutes for ~$8 (/showcase).

3. The 6-agent codebase audit

Six dimensions in parallel — 47 raw findings deduped to 12 unique, 4 of them P0, in about 9 minutes a pass (/showcase).

4. The pre-flight dependency swarm

One research agent per external dependency before any code gets written: 8 dependencies checked, 6 show-stoppers surfaced — including a Reddit API quote of $0.24 per 1K calls that re-priced the whole build (/showcase).

Orchestration patterns — waves, verification passes, when a swarm beats a single agent — are the job of /swarms; the self-orchestrating version is /dynamic-workflows, which has run 40 minutes unattended within its published limits (16 concurrent, 1,000 agents per run).

Which agents run while you sleep?

5. The 4 AM browser pricing watcher

A Playwright agent walks competitor pricing pages nightly — one 4:11 AM run caught a $79→$89 move and posted a 47-word Slack summary, in under 30 seconds of runtime (Ch 33).

6. The Codex night shift

A second vendor's agent on a loop against Sentry, PostHog, and BetterStack signals — it opened a PostHog funnel-fix PR in an isolated worktree overnight, and once built a desktop pet in ten minutes as a side effect (Ch 42).

7. The headless Sentry-fix cron

Headless Claude Code on a timer: read the newest Sentry issue, reproduce, patch, open a PR for human review (Ch 18). The 7:30 AM morning brief that summarizes all of it into one message is its sibling (Ch 7).

Can an agent talk — or write — as you?

8. The voice agent

$0.06–$0.40 per minute depending on the stack; a 90-second tier-1 support resolution at $0.40 against a $14 human-handled ticket. The failure that taught the most: a 1.4-second silence on a cold call — cold start 2.6s versus 900ms warm — long enough for the caller to say "hello?" (Ch 27).

9. The write-on-behalf persona agent

Drafts outbound in the operator's voice behind an approval gate: 60 drafts a week, 70% approved on first pass, median 6–9 minutes from signal to draft — numbers as published in Ch 34. The gate exists because one unsendable post triggered a co-founder phone call in four minutes.

Can an agent ship a whole product?

10. The one-shot native iOS app

Claude Code one-shot a native SwiftUI language-partner app from a single prompt. What it couldn't do: cross a single App Store gate — signing, the $99/year membership, review. The proof of the whole story is a $7.99 subscription renewal row, not a download count (Ch 45).

11. The $81 Saturday build

A working voice-brief product in a weekend for ~$81 in tokens, against a $2,000–8,000 contractor quote for the same scope (Ch 19).

What about an agent that manages its own memory?

12. The Dreaming pipeline

A weekly agent that digests 483 session transcripts (995 all-time), compresses 5.8MB of logs to a ~24K-token digest — about 16% signal — and proposes memory updates it is structurally forbidden to write itself. 15 of 15 proposals verified in the first reviewed cycle; the tool still missed 2 duplicates, which is why the human stays in the loop (Ch 44).

What do these agents cost, side by side?

Costs as of 2026-06; every row traces to its linked chapter.

Where do AI agents fail?

The other half of every receipt above — the section a vendor page will never publish:

• The browser watcher posted a competitor pricing screenshot into the wrong Slack channel — a customer saw another company's pricing (Ch 33).
• The same /goal harness that cleared a migration for $3.16 burned $11 across 41 turns when "done" couldn't be measured (Ch 38).
• The voice agent's 1.4-second silence on a cold start — the caller heard dead air on an investor call (Ch 27).
• The persona agent produced an unsendable post that triggered a co-founder phone call in four minutes — the approval gate is the product now (Ch 34).
• The iOS one-shot hit the signing maze: plausible-looking, wrong fixes for provisioning errors — the one part of shipping the agent couldn't brute-force (Ch 45).

FAQ

What is a real example of an AI agent?

A 15-agent swarm that wrote this 25,000-word book in ~6 minutes for under $40 — dispatched in one message, each agent writing one chapter to a file.

What do AI agents cost to run?

On this page's receipts: $3.16 for an 18-minute migration, $6.20 for a 20-document strategic plan, under $40 for a book, ~$81 for a shipped product — and $11 wasted on one unmeasurable goal.

Can AI agents run unattended?

Yes, with gates. The 4 AM watcher and the Codex night shift both run unsupervised — and the wrong-Slack-channel incident is why every unattended agent here posts to a human-reviewed surface first.

Can an AI agent build an entire app?

The code, yes — one shot. The platform gates (signing, $99 membership, review), no. The operator owns the gates; that's the honest division of labor.

What's the difference between an AI agent and a multi-agent swarm?

One agent is one worker with tools. A swarm is parallel agents in waves — 3–4 per wave, one dispatch message, a verification pass between waves.

Related: swarm orchestration · dynamic workflows · the live setup snapshot · what is agentic coding