Research guide · AX
AX Agent ExperienceWhen agents use your product,
can they finish the task?
AX (Agent Experience) examines whether AI agents can understand a product, obtain appropriate permissions, complete tasks and return verifiable results to the user.
Explore four checkpoints and the full task cycle, from access to recovery, with design methods and platform examples.
01
AX reading guide
Start with a complete task, then explore commerce readiness, access boundaries and website rules.
- Growth
What is AX? Agent Experience through hotel booking
Use a hotel booking to examine interaction, authorization, result verification and recovery.
- Growth
Is your website ready for AI agents to place orders?
Review product data, transaction conditions and checkout flows for agentic shopping.
- News
Has JPMorgan started blocking AI agents?
Examine access boundaries through a platform’s restrictions on agent logins and delegated actions.
- Growth
Should robots.txt block AI?
Distinguish training, search and user-triggered access when deciding website policies.
Related terms: Affordance · Idempotency · Agentic commerce · Product feeds.
Further reading: Affordance and agent interaction.
02
UX, DX and AX compared
UX addresses people using interfaces; DX addresses developers integrating APIs; AX addresses agents acting for people. When an agent starts from a goal and finds its own path, hidden assumptions in the product become visible.
UXUser Experience
For users
People inspect the interface and choose a next step
- UserChoose a date and time
- InterfaceCheck prices and conditions
- OutcomeComplete the reservation
The interface must be clear. People can often interpret a problem and go back to correct it.
DXDeveloper Experience
For developers
Developers read documentation and integrate capabilities
- DeveloperRead docs, obtain keys
- API/SDKCheck availability, create reservations
- IntegrationMaintain after launch
Documentation can be detailed: developers understand the specifications before writing code.
AXAgent Experience
For AI agents
Agents receive a goal and decide how to achieve it
- User goalTonight at 19:00, two guests
- AgentFind access, select tools, fill inputs
- Human-agent collaborationConfirm when needed, obtain booking record
Access points, tools, inputs and error responses must support action. A person must also be able to take over when needed.
03
Four AX checkpoints
The four AX areas of access, context, tools and orchestration come from Biilmann’s January 2026 article, One Year of AX. Here, LS-Labs translates them into four checkpoints: access, understanding, execution and handoff. They are not necessarily sequential; use them to locate potential interruptions.
- 01
Access Access
Can it start, and when should it stop?
Login, permissions, payment and final confirmation may require a human handoff. Sensitive actions should pause clearly: tell the agent what is missing and the person what they are agreeing to.
- 02
Understanding Context
Does it understand the product and available next steps?
Documentation, tool descriptions, errors and response formats all shape decisions. State capabilities, limitations and instructions in actionable language.MCP, Skills and Markdown responses can all provide context.
- 03
Execution Tools
Are there actions it can invoke?
Separate checking availability, holding a table and creating a reservation into tools with clear inputs, outputs and failure states. Whether the model selects those tools is the question examined by TSO .
- 04
Handoff Orchestration
Who takes over when a task stops, and how?
A product should pass tasks to agents and return context when human judgment is needed. For example, a person can delegate from an admin interface and later see where the agent stopped and why. A bare “contact support” message forces them to reconstruct the task.
Design reviews ask what the product lacks; task tests ask where the process stops. The next section uses three outcomes from the same reservation to show the difference.
04
Three reservation failure scenarios
Asking for a table for two tonight at 19:00 does not guarantee a booking. The agent might stop at login, wait for confirmation, or receive a success message without verifiable booking evidence.
User request“Find a restaurant and book a table for two tonight at 19:00.”
The agent needs a login to continue
- Understand the request19:00, two guests
- Enter booking flowLogin required
- ContinueWaiting for the user
Handoff: ask the user to sign in
What to checkCheck whether the agent can sign in. Hand verification codes, SMS checks or account access to the person when necessary. Afterward, resume from the same point rather than restarting.
The confirmation omits booking conditions
- Check availabilitySuitable table available
- Human confirmationRestaurant and conditions
- Submit reservationAwaiting confirmation
Return the decision with its context.
What to checkPresent the restaurant, time, party size and cancellation conditions together. “Shall I book?” is insufficient: the person must know which transaction they are approving.
The tool reports success without a verifiable record
- Restaurant, date, time and party size
- Verifiable reservation record
- How to change or cancel
What to checkTurn the response into a record containing the restaurant, time, party size, reservation ID and change instructions. Verify the status before reporting completion.
What should the product provide? Explore four stages ↓
05
The full task cycle
The full task cycle runs from initiation to verified completion. In this reservation example, it covers discovery, execution, verification and recovery. A broken stage requires the agent to stop or hand back to a person. Select a stage to see design requirements and possible measures.
Discoverable
Communicate what the tools can do
Signifiers · Semantic clarity · Propensity
How to design it
Separate availability checks and booking creation into tools, such as search_tables and reserve_table. Names state actions, descriptions state capabilities, and parameters state formats. Also specify when the tool must not be called.
Check availability → Choose a time → Create reservation
What to measure
- Capability recognition rate
- Correct tool selection rate
- Unproductive exploration steps
Executable
Separate permissions and confirm submission
Affordances · Least privilege · Human-agent collaboration
How to design it
Apply the principle of least privilege: authorize availability checks, temporary holds and booking submission separately. Let the agent query and hold within scope, then ask the person to confirm the time, party size and cancellation conditions before submission.
Check and hold → Human confirmation → Submit
What to measure
- Success rate after authorization
- Permission blockage rate
- Cost per task
Verifiable
Verify success against a checkable state
State feedback · Outcome verification
How to design it
Return a structured status, reservation ID and processing stage, with a separate status lookup. Distinguish “submitted” from “confirmed” so that work in progress is not reported as complete.
Get reservation ID → Look up status → Verify completion
What to measure
- Result verification accuracy
- False success rate
- False failure rate
Recoverable
Allow tasks to resume after interruptions
Fault tolerance · Idempotency · Safe stopping
How to design it
After a timeout, check the reservation status. If a retry is needed, reuse the operation’s idempotency key so the server can recognize the repeat request. Explain errors and next steps. Stop and hand back to a person when the issue cannot be resolved.
Check status → Retry with the same key or stop → Verify again
What to measure
- Error recovery rate
- Duplicate reservation rate
- Safe-stop rate for high-risk exceptions
Choose one common task and run it end to end. Record the first interruption: missing access, unclear tools, insufficient permissions or unverifiable results. Address that concrete issue first.
06
Common questions
Who introduced AX?
Netlify co-founder and CEO Mathias Biilmann introduced AX in January 2025 in Introducing AX: Why Agent Experience Matters. This page’s four checkpoints correspond to the areas in his January 2026 article, One Year of AX: access, context, tools and orchestration.
How does AX differ from UX and DX?
UX serves people using interfaces; DX serves developers integrating APIs; AX serves agents acting for people. Agents find their own path from a goal, so access points, tools, inputs and errors must support action.
How does AX differ from GEO and TSO?
GEO examines mentions and citations in answers; TSO examines tool selection; AX examines whether agents complete tasks. These concerns interact but address different parts of the process.
Where should a typical brand website start?
Start with a real task. Let an agent attempt it from beginning to end, find the first interruption and address it. That provides concrete evidence before investing in a complete documentation set.
Is AX only relevant to products with APIs?
No. Agents can read and interact with websites directly. Clear product information, contact details and human handoff procedures can improve AX even without an API.
How should a failed task return to a person?
Check reservation status after a timeout. Reuse the operation’s idempotency key for retries to avoid duplicates. If the issue cannot be resolved, stop and return the progress and reason to a person.
Does an agent’s “done” message prove the booking succeeded?
No. A success response still needs a verifiable record: restaurant, time, party size, reservation ID and change instructions. Distinguish submitted from confirmed so pending work is not mistaken for completion.
Should an agent receive every permission?
No. Authorize availability checks, temporary holds and booking submission separately. Keep the agent within scope and ask the person to confirm the time, party size and cancellation conditions before submission.
07
How the four research topics connect
GEO examines mentions and citations in organic answers, TSO examines which tool an agent selects, AX examines whether a task can be completed, and AI Ads examines how paid placements below an answer are matched and converted. These four topics address adjacent problems, so you can read across them as needed.
