Skip to content

Agentic Empathy

Agentic empathy means reducing uncertainty for constrained, tool-using agents that fetch, parse, compare, cite, and sometimes act on behalf of a human.

Use this page when you need the source-backed expansion behind the Agentic Empathy slide in the Codeable Skill Chat deck.

This is not anthropomorphism. The useful move is to respect agent constraints: entry path, context budget, fetch loss, provenance, and action boundaries.

Agentic empathy is the habit of asking:

If an agent lands here with a task, a tool budget, and an imperfect fetch pipeline, what can it find, understand, trust, cite, and safely do?

The Codeable deck now treats this as a two-step explanation:

  1. What it is and why it matters: design for what an agent can actually perceive, preserve, verify, and safely do on a human’s behalf.
  2. What it looks like: source-backed examples of structure, fetch quality, crawler lanes, and bounded actions.

Agentic empathy is not anthropomorphism. It is a practical design habit: respect the agent’s entry path, context budget, fetch loss, source uncertainty, and action risk.

That makes it close to familiar UX empathy, but aimed at a different immediate reader. Human UX empathy asks what the visitor is trying to do and where the experience creates friction. Agentic empathy asks what a constrained, tool-using agent can perceive and carry forward while acting for a human.

Agents increasingly mediate discovery, citation, comparison, and actions. If a site’s truth does not survive fetch, parsing, citation, and action boundaries, the agent may guess, skip the site, cite a competitor, or follow an unsafe affordance.

Microsoft’s Design Foundations for Agents is useful general support for this framing: trusted agent experiences require more than visual polish, including reliability, safety, privacy, transparency, accountability, human-centered scope, and error recovery.

This page expands the source-linked examples from the slide deck.

ExampleWhat it teachesWhy it matters
Vercel Agent ReadabilityAgent-readable pages need discoverable structure, metadata, context, Markdown-friendly content, and validation.Structure is empathy: it gives agents a predictable path through the page.
Cloudflare AI consumability and Markdown for AgentsA large docs site can expose index.md, llms.txt, llms-full.txt, page-level Markdown, and content negotiation.Markdown and source maps are useful affordances, but they should point back to maintained source content.
Agent-Friendly Documentation SpecCoding agents can lose content through truncation, tabbed docs, redirects, soft 404s, auth gates, broken Markdown, and oversized pages.Fetch loss is real. If the important answer disappears in serialization, the agent may never see it.
OpenAI crawlers, Anthropic crawler controls, and Perplexity crawlersSearch/index crawlers, user-directed fetchers, and training/model-development crawlers are separate policy lanes.Entry path matters. “Block AI bots” is too blunt when the site still wants AI-search or user-fetch visibility.
Microsoft agentic risk guidance and Simon Willison on the lethal trifectaTool-using agents need boundaries around permissions, approvals, logs, private data, untrusted content, and external communication.Actions need boundaries. A capability signal should describe something real, scoped, and monitored.
Agent questionWebsite signal
Can I find it?Sitemap, internal links, llms.txt, and search crawler access.
Can I fetch it?Clean 200 responses, low redirects, WAF allowlists, and user-fetch bot access.
Can I parse it?Semantic HTML, clear headings, low boilerplate, and Markdown alternatives.
Can I trust it?Dates, authorship, visible sources, methodology, and schema parity.
Can I cite it?Short answer passages, named entities, stable URLs, and visible facts.
Can I act safely?Real APIs or tools, explicit auth, scoped permissions, approvals, and logs.
  • “Empathy here does not mean the agent has emotions. It means we respect its constraints.”
  • “The agent may be reading a compressed, converted, truncated version of the page.”
  • “A pretty accordion can be a hostile document if the answer disappears in serialization.”
  • “Dates, sources, and direct answer passages are not SEO tricks. They are how an agent knows what it can responsibly quote.”
  • “Do not make fake doors for agents. If a manifest says there is a tool, auth flow, or API, it needs to be real.”