Building Software Products for the Agentic Age

CLI · MCP · Agentic web

Your next user reads text, calls tools, and clicks as a last resort.

Indrajeet Patil

Source code for these slides can be found on GitHub.

What you’ll learn today

📋 We assume you know

• what a large language model (LLM) is
• an agent is an LLM calling tools in a loop
• CLI, HTTP, and HTML basics

🎯 You will leave with

• how agents experience products
• products agents act on
• websites agents read and cite
• tools to measure readiness

Agent-friendliness is a new layer on top of good engineering, not a replacement for it.

A new kind of user

Part 0 · Foundations

Same product, new user

For decades, every user reached your product through a screen. Now many arrive through an agent.

Agents can see and click, but don’t require it. Offer text interfaces and a path for human approval.

How an agent works with your product

Text works best. Output, errors, and docs all land in the context window.

Every token costs. Noisy output burns money and crowds out the task.

No one at the keyboard. A y/N prompt stalls the loop; approval must come through the host.

Humans and agents need different things

Need	Human user	Agent user
Input	Pixels on a screen	Text in a context window
Navigation	Menus, clicks, scrolling	Commands, links, tool calls
Learning	Tutorials, tooltips	--help, schemas, llms.txt
Errors	Red banner, retry button	Exit code + message it can act on
Login	Form, password, MFA prompt	Token, OAuth flow, environment variable
Noise	Mildly annoying	Costs tokens and derails the task

Good news: most of what agents need is what good engineering already recommends.

Products agents can operate

Part 1 · CLI and MCP

Three ways in

	HTTP API	CLI	MCP server
Agent needs	An HTTP client and your docs	A shell	An MCP-capable host
Discovery	OpenAPI spec, docs pages	--help, man pages, training data	server/discover, then tools/list returns schemas
Auth	API keys, OAuth	Token in env var, stored login	Per-user OAuth (remote) or local credentials
Context cost	Whatever docs it reads	Pay only for commands used	Tool schemas, up front or searched on demand (host’s choice)
Sweet spot	Everything builds on it	Coding agents with a terminal	Chat apps and hosted agents with no shell

CLI and MCP are thin layers over a good API. Get the API right first.

Hosts can defer MCP tool schemas and load them on demand: see Anthropic’s tool search tool.

The command line is back

DEC VT100 terminal. Photo: Jason Scott, CC BY 2.0

Fluent

Models have read decades of shell scripts and docs.

Composable

Pipe into jq, grep, and xargs.

Self-documenting

--help is read only when needed.

Cheap

Nothing enters context until it runs.

Anatomy of an agent-friendly CLI

Principle	Why agents care	Looks like
Structured output	Parse, don’t scrape tables	--json, -o json
Field selection	Smaller output, fewer tokens	--json name,id, --query
Never block on input	No one is there to type y	--yes, GH_PROMPT_DISABLED=1
Headless auth	Agents can’t click “Sign in”	GH_TOKEN=…, az login --identity (on Azure)
Meaningful exit codes	Branch on success vs failure	0 ok, 2 bad usage, 3 not found, 4 refused
Actionable errors on stderr	Self-correct on the next turn	error: no todo 7; run todo list
Examples in --help	Learn usage on demand	gh pr create --help
Stable, versioned output	Scripts and agents break silently	Add fields; never rename them

Sources: Arcjet, Speakeasy, GitHub CLI manual

A minimal agent-friendly CLI

A toy todo CLI in Typer. Highlights mark the agent-friendly bits.

todo.py

from json import dumps
import typer

app = typer.Typer(rich_markup_mode=None)  # plain text: no box art for agents
TODOS = [{"id": 1, "title": "Write slides"}]

@app.command("list")
def list_todos(json: bool = typer.Option(False, "--json")):
    """Show all todos."""
    print(dumps(TODOS) if json else "\n".join(t["title"] for t in TODOS))

@app.command()
def delete(todo_id: int, yes: bool = typer.Option(False, "--yes")):
    """Delete a todo by id."""
    if todo_id not in {t["id"] for t in TODOS}:
        typer.echo(f"error: no todo {todo_id}; run `todo list`", err=True)
        raise typer.Exit(3)
    if not yes:
        typer.echo("error: deleting needs --yes", err=True)
        raise typer.Exit(4)
    print(dumps({"deleted": todo_id}))

app()

A stub: delete validates and reports but stores nothing. Typer turns type hints and docstrings into --help. Exit code 2 already means a usage error, so domain errors start at 3. Tested with Typer 0.27.3.

The same CLI, seen by an agent

Human mode

terminal

$ todo list
Write slides

Agent mode

terminal

$ todo list --json
[{"id": 1, "title": "Write slides"}]

Failures that teach the agent

terminal

$ todo delete 7; echo "exit $?"
error: no todo 7; run `todo list`
exit 3

$ todo delete 1; echo "exit $?"
error: deleting needs --yes
exit 4

$ todo delete 1 --yes
{"deleted": 1}

Each error tells the agent what went wrong and what to try next.

Exemplar: GitHub CLI (gh)

Screenshot of cli.github.com, 7 Oct 2026

• --json + --jq almost everywhere
• GH_TOKEN for headless auth
• GH_PROMPT_DISABLED stops prompts

terminal

# Ask for exactly the fields you need
$ gh pr list --repo cli/cli --limit 3 \
    --json number,author \
    --jq '.[] | "\(.number) \(.author.login)"'
14620 williammartin
14617 bingtang9
14583 BagToad

# Escape hatch: any REST endpoint, authenticated
$ gh api repos/cli/cli --jq '.license.spdx_id'
MIT

Source: GitHub CLI manual

Exemplar: Azure CLI (az)

terminal

# JMESPath query + TSV output: values only, no table art
$ az vm list --show-details \
    --query "[?powerState=='VM running'].name" -o tsv

# No prompt; --no-wait returns once started
$ az group delete --name rg-demo --yes --no-wait
$ az group wait --name rg-demo --deleted

# Managed identity: on an Azure VM or app
$ az login --identity

# Escape hatch: any Azure REST API, already authenticated
$ az rest --method get --url "<management-endpoint>"

• --query filters before output
• -o json, tsv, or table everywhere
• --yes, --no-wait for unattended runs, then wait to confirm

Pattern: human-friendly by default, machine-friendly by flag, with the full API one command away.

Sources: Azure CLI: query output with JMESPath, az group, managed identity sign-in, Azure MCP Server overview

What is MCP?

The Model Context Protocol is an open standard for plugging tools and data into any agent host. Current spec: 2026-07-28

%%{init: {"htmlLabels": false, "flowchart": {"htmlLabels": false, "padding": 16}, "themeVariables": {"fontFamily": "Space Grotesk, sans-serif", "fontSize": "20px"}}}%%
flowchart LR
    subgraph Host["Agent host"]
        M[LLM] <--> C[MCP client]
    end
    C <-- "server/discover<br/>tools/list<br/>tools/call" --> S[Your MCP server]
    S <--> API[Your API]

    classDef box fill:#ffffff,stroke:#0a0a0a,stroke-width:3px,color:#0a0a0a
    classDef yours fill:#ffd600,stroke:#0a0a0a,stroke-width:3px,color:#0a0a0a
    class M,C box
    class S,API yours
    style Host fill:#f5f5f0,stroke:#0a0a0a,stroke-width:3px,color:#0a0a0a

Tools: actions the model can call

Resources: data the app can read

Prompts: templates the user can pick

Source: MCP specification 2026-07-28

A minimal MCP server

server.py: you write

from mcp.server import MCPServer

mcp = MCPServer("Todo")


@mcp.tool()
def add_todo(
    title: str, due: str | None = None
) -> dict:
    """Create a todo. due: ISO date."""
    return {"id": 42, "title": title}


if __name__ == "__main__":
    mcp.run("streamable-http")

tools/list: the agent sees

{
  "name": "add_todo",
  "description": "Create a todo. due: ISO date.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "title": {"type": "string"},
      "due": {"anyOf": [{"type": "string"},
                        {"type": "null"}]}
    },
    "required": ["title"]
  }
}

Type hints and docstrings become the interface. Write them for a reader with no context.

A stub: add_todo stores nothing and ignores due. Tested with the MCP Python SDK 2.3.0, which speaks protocol 2026-07-28; schema output abridged.

Every request stands alone

POST /mcp: headers

Content-Type: application/json
Accept: application/json, text/event-stream
MCP-Protocol-Version: 2026-07-28
Mcp-Method: tools/call
Mcp-Name: add_todo

POST /mcp: body

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "add_todo",
    "arguments": {"title": "Write slides"},
    "_meta": {
      "io.modelcontextprotocol/protocolVersion":
        "2026-07-28",
      "io.modelcontextprotocol/clientCapabilities":
        {}
    }
  }
}

200 OK: the result

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "resultType": "complete",
    "content": [{
      "type": "text",
      "text": "{\"id\": 42, ...}"
    }],
    "isError": false
  }
}

No sessions, so any replica can take the request. Cache discovery and reads by their hints; keep app state in a shared store.

Sent with curl to the minimal server.py from the previous slide (MCP Python SDK 2.3.0); without the headers the server returns 400. Response abridged. Specs: Streamable HTTP, caching.

CLI or MCP?

Prefer a CLI when

• the agent has a terminal (coding agents)
• one developer, their own credentials
• output needs piping and scripting
• humans use the same tool

Prefer MCP when

• the agent has no shell (chat apps)
• many users need per-user OAuth
• you need central governance and audit
• you want hosted, zero-install access

Design beats protocol

A naive MCP server exposes one tool per endpoint. Cloudflare’s API has 2,500+ endpoints.

~1.17M tokens of tool definitions: one tool per endpoint

~1,000 tokens of tool definitions: search() + execute()

Expose discoverable, task-shaped tools, not raw endpoints. Let the agent search for detail on demand.

The numbers compare what the tool definitions cost in context, not total task tokens, latency, or success rate. Source: Cloudflare, Code Mode: give agents an entire API in 1,000 tokens

Who does this well?

Company	CLI	MCP server	Docs agents can read
GitHub	gh with --json, --jq, gh api	GitHub MCP Server, local and remote	Markdown docs
Microsoft Azure	az with --query, -o tsv, az rest	Azure MCP Server	Microsoft Learn MCP
Stripe	stripe CLI	Remote MCP at mcp.stripe.com	docs.stripe.com/llms.txt
Cloudflare	wrangler	Code Mode MCP for the whole API	Markdown for Agents, llms.txt
Vercel	vercel CLI	Remote MCP at mcp.vercel.com	Markdown via content negotiation

These vendors’ users span terminals and chat apps, so they ship all three. Choose yours from your users’ workflows.

Teach the agent: ship a skill

skills/todo-cli/SKILL.md

---
name: todo-cli
description: Manage todos with the `todo` CLI.
  Use when asked to list or delete todos.
---

# Using the todo CLI

- Run `todo list --json` and parse stdout.
- Exit code 3 means "not found":
  run `todo list` to get valid ids.
- Deleting needs `--yes`; ask the user first.

Agent Skills are folders of instructions an agent loads only when relevant.

AGENTS.md does the same job for a repository.

Tools ship them too: Astral publishes an official uv skill.

Sources: agentskills.io, agents.md, Astral’s uv skill

Packages: docs agents can fetch

Coding agents read your docs, in Markdown, before they call your code.

File	What it holds	uv
/llms.txt	Summary plus curated links to Markdown pages	5 KB
/llms-full.txt	Every docs page concatenated into one file	—
/<page>/index.md	A Markdown twin of each HTML page	29 KB vs 169 KB HTML
<link> tag	rel="alternate" points crawlers at the .md twin	—

llms.txt to navigate, .md pages to read one topic, llms-full.txt to load everything.

Hype check: Google says AI search needs no llms.txt, but it’s cheap and Lighthouse checks it. Start it with a “When to use this” line.

Sizes from uv’s docs, 7 Oct 2026; — means not published. /.well-known/ copies count only at the domain root (RFC 8615). Sources: llmstxt.org, Google, Lighthouse.

Generate them at build time

mkdocs.yml: you write

plugins:
  - llmstxt:
      markdown_description: |
        You can use uv to install Python
        dependencies, run scripts, ...
      sections:
        Getting started:
          - getting-started/installation.md
        Guides:
          - guides/scripts.md

llms.txt: the build writes

# uv

> uv is an extremely fast Python package
> and project manager, written in Rust.

You can use uv to install Python ...

## Getting started

- [Installation](https://docs.astral.sh/uv/
  getting-started/installation/index.md)

## Guides

- [Running scripts](https://docs.astral.sh/uv/
  guides/scripts/index.md)

The llmstxt plugin for MkDocs writes these files on every build. Zensical has it built in since 0.0.67.

Abridged from uv’s mkdocs.yml and llms.txt (long lines wrapped). Sources: mkdocs-llmstxt, Zensical plugins

Get indexed by Context7

Screenshot of context7.com/astral-sh/uv, 7 Oct 2026

Add your repo at context7.com and steer the parser with context7.json. Agents then call resolve-library-id and query-docs over MCP.

what the agent gets back

# query-docs("/astral-sh/uv",
#   "add a dependency to a project")
# → "Adding a dependency with uv add"
#   (docs/guides/projects.md)
uv add requests
uv add --dev pytest
uv add --group lint ruff

Sources: Context7, Adding libraries and Library owners. Output abridged from a live query on 7 Oct 2026.

Context7 as a docs quality check

4,273 code snippets extracted

637,996 tokens of docs indexed

80 benchmark score

Signal	What it tells you
Snippets	How many code examples Context7 extracted. Few snippets means few examples; extraction doesn’t prove they run.
Benchmark	LLMs ask the questions developers ask, answer from your docs, and a jury scores the answers.
Trust score	Organisation signals: age, stars, followers, contributors, referrals.

Ask Context7 your users’ top five questions. No answers? Your docs are hard to find there.

Source: Upstash, Inside Context7’s quality stack. uv figures from Context7 on 7 Oct 2026. Scores are LLM-judged signals, not a certificate of correctness or security.

Agents act with your users’ credentials

🔑 Least privilege

Fine-grained, scoped, short-lived tokens. Read-only by default.

🧪 Dry runs, safe retries

--dry-run first; idempotency keys so retries never write twice.

🛑 Explicit intent

A flag for anything destructive; consent comes from the user.

🧾 Audit trail

Identify agent traffic and log every action it takes.

Treat tool output as untrusted input: text in the data can carry instructions that hijack the agent (prompt injection).

Sources: MCP tools: user interaction, Stripe idempotent requests

Websites agents can read

Part 2 · The agentic web

The web was built for people with browsers

The first web server, CERN. Photo: Coolcaesar, CC BY-SA 3.0

< 50% of HTML page requests on Cloudflare now come from a human

Machines, from crawlers and scrapers to agents acting for people, now request more pages than people do.

Source: Cloudflare, From ranking to recommended (Aug 2026)

Who are the machine visitors?

Visitor	Example user agents	What it wants	When
Search crawler	Googlebot, OAI-SearchBot, Claude-SearchBot	Index pages for search results	Continuously
Training crawler	GPTBot, ClaudeBot, CCBot	Content for training models	Periodically
User-directed fetcher	ChatGPT-User, Claude-User, Perplexity-User	Answer one person’s question now	On demand
Browser agent	Agents driving Chrome or a headless browser	Click and type on a user’s behalf	On demand

Each group can be allowed or refused separately. That choice is yours.

Google-Extended is a robots.txt control token for Gemini training, not a separate crawler. Sources: OpenAI, Anthropic, Google

Participation is a choice

🚫 Opt out

robots.txt

User-agent: GPTBot
User-agent: ClaudeBot
User-agent: CCBot
User-agent: OAI-SearchBot
User-agent: Claude-SearchBot
Disallow: /

⚖️ Search, no training

robots.txt

User-agent: GPTBot
User-agent: ClaudeBot
User-agent: CCBot
Disallow: /

User-agent: *
Allow: /
Content-Signal: search=yes,
  ai-input=yes, ai-train=no

✅ All in

robots.txt

User-agent: *
Allow: /
Content-Signal: search=yes,
  ai-input=yes, ai-train=yes

Not taking part is completely legitimate. Pick deliberately, then make the signals match.

robots.txt is a request, not a lock: enforce with bot management or a WAF. Content signals state a preference; Disallow is what crawlers without signal support obey. Content-Signal values must sit on one line in a real file. Sources: contentsignals.org, Cloudflare managed robots.txt

The old rules still apply

Agent extras: llms.txt, Markdown, MCP

Structured data and metadata

SEO: crawlable, titled, described, in a sitemap

Fast, accessible, semantic HTML over HTTPS

“All existing SEO fundamentals continue to be worthwhile.”

Google Search Central

“Agents review the accessibility tree to identify interactive elements.”

Chrome Lighthouse docs

Sources: Google: AI features and your website, Chrome: accessibility for agents

A ladder, not a checklist

LEVEL 0 Foundations SEO, accessibility, performance, HTTPS

LEVEL 1 Discoverable robots.txt stance, sitemap, content signals

LEVEL 2 Readable Server-rendered HTML, Markdown, llms.txt

LEVEL 3 Trustworthy Structured data, trust pages, honest status codes

LEVEL 4 Operable API catalogue, OAuth, MCP, WebMCP

Climb only as high as your site needs. Personal sites can stop at level 3.

Level 1: be discoverable

robots.txt

User-agent: *
Allow: /
Content-Signal: search=yes, ai-input=yes, ai-train=no

# A signal alone won't stop training crawlers
User-agent: GPTBot
User-agent: ClaudeBot
User-agent: CCBot
Disallow: /

Sitemap: https://example.com/sitemap.xml

Crawlers obey only their most specific group, so named training bots skip * and see only Disallow.

robots.txt sits on 78% of top sites; content signals on just 4%.

Sources: Cloudflare, Introducing the Agent Readiness score (Apr 2026); Anthropic: blocking ClaudeBot

Level 2: put content in the HTML

❌ Client-rendered shell

<body>
  <div id="root"></div>
  <script src="/app.js"></script>
</body>

Many fetchers never run JavaScript, so they see nothing.

✅ Server-rendered page

<body>
  <main>
    <h1>Pricing</h1>
    <p>Pro plan: $20 per month.</p>
  </main>
</body>

The answer is in the first response.

terminal

# What does an agent see? Expect a 200 and the prices in the HTML.
curl -iL -A 'Claude-User/1.0' https://vercel.com/pricing

Source: Vercel, Make your site readable by AI agents

Level 2: serve Markdown to agents

terminal

$ curl -sI -H 'Accept: text/markdown' \
    https://vercel.com/blog/making-agent-friendly-pages-with-content-negotiation
content-type: text/markdown; charset=utf-8
vary: Accept

645 → 7 KB Vercel blog post, HTML vs Markdown

179 → 17 KB Cloudflare docs page, HTML vs Markdown

Can’t negotiate? Publish .md mirrors and link them with <link rel="alternate" type="text/markdown">.

Measured with curl on 7 Oct 2026. See Vercel, content negotiation, and Cloudflare, Markdown for Agents.

Level 3: say who you are, in data

index.html

<html lang="en-GB">
<head>
  <title>Jane Doe, data engineer</title>
  <meta name="description" content="...">
  <link rel="canonical" href="https://jane.dev/">
  <script type="application/ld+json">
  {
    "@context": "https://schema.org",
    "@type": "ProfilePage",
    "mainEntity": {
      "@type": "Person",
      "name": "Jane Doe",
      "sameAs": ["https://github.com/janedoe"]
    }
  }
  </script>
</head>

Signal	Keep it
<title>, description	Unique per page
Canonical URL	Exactly what is served
lang, og:locale	Consistent everywhere
JSON-LD type	Honest: no fake FAQs
Dates	Real changes only

Source: schema.org. Mismatched locale or canonical signals confuse crawlers and retrieval alike.

Level 3: earn the agent’s trust

🪪 Trust pages

Real About, Contact, and Privacy pages, with substance, linked from every footer.

🔢 Honest status codes

404 for missing pages, not a 200 app shell. 429 with Retry-After when throttling.

🧭 Helpful 404s

Point lost agents to the sitemap, llms.txt, and the About page.

🔒 Accurate privacy

Describe the analytics you actually run. Agents quote you.

Sources: Vercel, Make your site readable by AI agents; is-agentic.com methodology

Level 4: advertise what agents can do

Signal	Location	Publish only if…
API catalogue (RFC 9727)	/.well-known/api-catalog	You have a public API
OpenAPI spec	e.g. /openapi.json	You have a public API
OAuth resource metadata (RFC 9728)	/.well-known/oauth-protected-resource + resource path	Your API or MCP server needs a token
MCP Server Card	/.well-known/mcp/server-card.json	You run an MCP server
Agent Skills index	/.well-known/agent-skills/index.json	You have tasks worth teaching
Agentic commerce	x402, ACP, UCP	Agents can buy from you

Never advertise a capability you don’t have. A false signal is worse than none.

For a server at /mcp, that is /.well-known/oauth-protected-resource/mcp; WWW-Authenticate can also point to it. The metadata names your authorisation server; that server, often an external identity provider, publishes /.well-known/oauth-authorization-server. Sources: Cloudflare, isitagentready.com check list; MCP authorisation

Level 4: WebMCP, tools inside the page

search.html

<form toolname="search_flights"
      tooldescription="Search flights by route and date">
  <input name="origin" required>
  <input name="destination" required>
  <input name="date" type="date" required>
  <button>Search</button>
</form>

Experimental

A browser agent calls search_flights directly instead of guessing which box to type in.

WebMCP is a Chrome origin trial; API names are still changing. Watch, don’t bet the roadmap.

Source: Chrome, WebMCP schema validity audit

Your host decides some of this

Capability	Static host (e.g. GitHub Pages)	Programmable edge (Cloudflare, Vercel, Netlify)
robots.txt, sitemap, llms.txt	✅	✅
.md mirrors + <link rel="alternate">	✅	✅
Link response headers (RFC 8288)	❌ fixed headers	✅
Accept: text/markdown negotiation	❌	✅ toggle or rewrite rule
Bot management, enforced blocking	❌	✅

Some failing checks are hosting decisions, not code bugs. Use the static substitutes and move on.

Personal vs commercial sites

	Personal, non-commercial	Commercial
Goal	Be found and cited correctly	Be found, recommended, and used
Must	Levels 0–1, server-rendered HTML	Levels 0–3 everywhere
Worth it	llms.txt, .md mirrors, Person JSON-LD, trust pages	Markdown negotiation, OpenAPI, CLI, MCP, OAuth
Skip	API catalogue, OAuth, MCP card, WebMCP: nothing to call	Signals for features you don’t offer
Track	Search Console, an occasional scan	Citation rate and share of voice in AI answers

Size the effort to the site. If visitors can’t do anything there, skip the API layer.

Source for answer-engine metrics: Cloudflare, From ranking to recommended

Measure and iterate

Part 3 · Tools

Paste a URL, get a score

isitagentready.com by Cloudflare, 7 Oct 2026

Discoverability, content, bot access, protocol discovery, and (unscored) commerce

is-agentic.com by Vercel, 7 Oct 2026

100+ checks, scored 0–100, with a web UI, CLI, API and MCP server

Same site, two very different scores

stripe.com on isitagentready.com, 7 Oct 2026

stripe.com on is-agentic.com, 7 Oct 2026

27 vs 71. Scanners are opinionated checklists, so read the findings, not the number.

From findings to a prompt

isitagentready.com: “Improve the score”

is-agentic.com: “Prompt to improve”

Both tools turn failures into a prompt for your coding agent. Review the diff like any other PR.

Scan from the terminal and CI

terminal

# Vercel's Is Agentic scanner, from the command line
npx is-agentic example.com --json

# Or read a stored report over HTTP
curl -s 'https://is-agentic.com/api/v1/report?url=example.com'

# Catch regressions in CI: set the threshold from your own baseline
npx @vercel/agent-readability audit https://example.com --min-score 80

Chrome Lighthouse has an experimental Agentic Browsing category: accessibility tree, layout shift, llms.txt, WebMCP.

The scanners practise what they preach: they offer CLIs, APIs, and MCP servers of their own.

Scores shift with methodology and certify nothing (Is Agentic’s limitations). Sources: Vercel KB, @vercel/agent-readability, Lighthouse agentic browsing

Most of the web isn’t there yet

Adoption among Cloudflare Radar’s top 200,000 domains, Apr 2026:

robots.txt present78% Content Signals4% Markdown negotiation3.9% MCP Server Card<15 sites

Most sites stop at robots.txt. A few small changes already put you ahead.

Source: Cloudflare, Introducing the Agent Readiness score

Wrapping up

Recap

Where to start

When	Product (CLI, MCP)	Website
This week	Add --json and non-interactive flags; fix exit codes	Decide your AI stance; update robots.txt and the sitemap
This month	Headless auth with scoped tokens; llms.txt for your docs	Server-render key pages; add llms.txt and .md mirrors
This quarter	Ship a skill; add MCP if agents without a shell need you; get indexed by Context7	Structured data, trust pages, Markdown negotiation
Always	Version the output; never rename fields	Keep SEO, accessibility, and performance healthy

Test real agent tasks before and after: success, wrong actions, time, and tokens. Use scanner findings to diagnose.

Key takeaways

1. Agents read text best and have finite memory. Design for text and tokens.

2. A great API, plus the CLI or MCP your users’ agents need, lets them operate your product.

3. For the web, SEO and accessibility still come first. Agent extras build on them.

4. Opting out is fine. Whatever you choose, send honest, consistent signals.

Build for humans and agents: the overlap is just good engineering.

Thank You

And happy building! 🤖

Check out my other slide decks on software development best practices

     

Resources

Vercel

• Is Agentic: paste a URL, get a score and a fix prompt
• Make your site readable by AI agents
• Agent-friendly pages with content negotiation
• Markdown and agent discovery

Cloudflare

• Is It Agent Ready?: scan and copy fix instructions
• Introducing the Agent Readiness score
• Markdown for Agents
• From ranking to recommended (AEO)
• Code Mode MCP

Standards

• Model Context Protocol
• llms.txt · Content Signals
• Agent Skills · AGENTS.md
• Context7: adding libraries

Further reading

• Google: AI features and your website
• Lighthouse agentic browsing audits
• Arcjet: designing a CLI for AI agents

Image credits

Slide	Image	Credit and licence
The command line is back	DEC VT100 terminal	Jason Scott, CC BY 2.0, resized
The web was built for people	First web server at CERN	Coolcaesar, CC BY-SA 3.0, resized
Exemplar: GitHub CLI	cli.github.com	Screenshot, GitHub, 7 Oct 2026
Get indexed by Context7	context7.com	Screenshot, Context7 (Upstash), 7 Oct 2026
Measure and iterate	isitagentready.com	Screenshots, Cloudflare, 7 Oct 2026
Measure and iterate	is-agentic.com	Screenshots, Vercel, 7 Oct 2026

Screenshots are used for commentary and remain the property of their owners. Visual design inspired by loot-drop.io.