How might we let a person say what an agent may achieve, not only which tools it may call?
Approach
The person writes a short task contract: an allow list, a deny list, a cost cap, an expiry time, and an approval list. A gateway checks every tool call against it.
Who pays
The platform team or security team of a company that lets agents act. It pays per agent, per month (price is an assumption).
Model
A fee per agent per month, inside an agent host or gateway product. Price is an assumption.
Enforced by
The agent host and the MCP gateway check each tool call against a signed task contract. The model vendor or the identity provider issues the contract.
How to try it
Open the Prototype tab. Change a tool from one list to another in the contract form. Press the tool buttons next to the chat. Read the gateway notice in the chat, then the panel and the timeline. Press Play on the diagram to replay the last call. Speed up the clock. See what happens after the contract expires.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
190,769 tokens measured. About $0.53 estimated. The Context tab lists the prompts.
How might we let a person see, limit, and end what an agent can do for them?
Approach
Each grant has a task, a scope, and an end time. A grant ends by itself. Every use writes a receipt.
Who pays
Compliance and GRC teams, and privacy officers (per workspace; assumption), at companies that hold customer data.
Model
A monthly fee for each workspace that uses the receipt feed and the audit export. Any price is an assumption. No price data exists yet.
Enforced by
The resource server writes a receipt for every use. The identity provider ends grants when the task ends. Auditors read the receipt feed.
How to try it
Open the Prototype tab. Press Ask for access. Read the agent request in the chat. Press a Grant button, or press Deny. Watch the receipt stream and the Receipt JSON panel. Set the clock speed to 60x. Press Revoke on a grant, or wait for it to end. Press Agents try again. Read the Denied receipt. Search the receipts. Filter by agent and by status. Press Show receipts as JSON. Copy the text from the box. Press Play under How it works. Read the five steps.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
203,100 tokens measured. About $0.57 estimated. The Context tab lists the prompts.
How might we hand work to a second agent without giving it more power than the first agent had?
Approach
Each handoff may add limits. It may never remove one. Anyone who holds the root key can check the whole chain.
Who pays
Agent framework vendors, and enterprise platform teams that chain agents.
Model
A fee per delegated chain, or per seat. Both are assumptions. This page sets no price.
Enforced by
Three parties act. Each hop's token may only add limits. The receiving tool checks the whole chain, like a macaroon. The orchestrator refuses a widened token.
How to try it
Open the Prototype tab. Press Play, or Step, to watch a hand-off chain from A to B to C. Press Start over. Hand off from A to B. Pick scopes, a cap, and a time. Ask for more than A has. Read the reason it fails. Hand off from B to C. Press Verify chain. Press Tamper. Press Verify chain again.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
186,075 tokens measured. About $0.52 estimated. The Context tab lists the prompts.
How might we stop a credential from leaving the place it was meant for when one agent calls another?
Approach
Put a firewall at the agent gateway. Bind each credential to one authority. Check header names, redirects, and destinations before any request leaves.
How to try it
Open the Prototype tab. Pick a scenario and a variant. Press Play in "Without firewall" mode. Watch the dot, read the chat, and read the red verdict. Press Play in "With firewall" mode. Read the firewall notice and the green verdict. Use Step to move one message at a time. Edit the rules and play again. Watch the "What the firewall sees" panel.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
190,576 tokens measured. About $0.53 estimated. The Context tab lists the prompts.
How might we make everything that trusts a broken key, token, or agent stop trusting it at the same time?
Approach
An issuer publishes one signed revocation notice to a shared feed. Every app that subscribes must act on it. The page shows who has not acted yet.
Who pays
Identity providers and SaaS security vendors. Large companies that run many connected apps also buy.
Model
A fee per connected app (assumption). No price is set.
Enforced by
The issuer signs and publishes each notice, and each consumer must poll or subscribe and drop the grant. A CDN can cache the feed and fan it out.
How to try it
Open the Prototype tab. Press Step with "Today" selected. Read each alert. Select "With a shared feed" and press Play. Compare the worst exposure window in both modes. Add a consumer that ignores the feed and press Play again.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
196,175 tokens measured. About $0.55 estimated. The Context tab lists the prompts.
How might we stop one stolen vendor credential from opening thousands of customer tenants?
Approach
Every grant names a purpose class, for example "backup read" or "ticket triage". The resource server checks the purpose on each use. A backup tool cannot assign roles.
Who pays
Managed service providers and cyber insurers who offer a premium credit. This is an assumption.
Model
An insurer premium credit for tenants that turn on purpose checks (assumption). A security add-on, priced per tenant per month (assumption, no price found).
Enforced by
The identity provider rejects a call outside the purpose class. A security vendor audits grants against a purpose catalog. The insurer asks for that audit report.
How to try it
Open the Prototype tab. Press Vendor credential stolen with all defenses off. Read the count. Turn on defense 1 only. Press the button again. Turn on defense 2. Read the server reply in the chat and the panel next to it. Turn on defense 3. Read the timeline for the held requests. Press Play on the diagram.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
119,913 tokens measured. About $0.34 estimated. The Context tab lists the prompts.
How might we make a leaked agent token useless to whoever finds it?
Approach
Bind each token to one audience and one sending key. Tag it "do not log" so scanners and gateways reject copies.
Who pays
Identity providers and AI gateway vendors, bought by their security product owners.
Model
A fee per agent identity inside an identity product, or a gateway add-on. Any price is an assumption: low cents per agent per month.
Enforced by
The identity provider issues only audience-bound and sender-bound tokens to agent clients. The gateway rejects tokens with no binding. The log pipeline drops tagged fields. A CI test replays a token and checks that it fails.
How to try it
Open the Prototype tab. Pick zero, one, two, or three defenses. Press "Leak the token". Press "Replay it". Read the event log and the "What the gateway sees" panel.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
105,911 tokens measured. About $0.30 estimated. The Context tab lists the prompts.
How might we show where content came from when the proof is often missing?
Approach
Use three states, not two: signed, unsigned, and stripped. Each state says why. Unknown never reads as fake.
How to try it
Open the Prototype tab. Pick an item in the feed and article views. Compare the common badge with the honest label. Select Play in the photo journey. Find the step that removes the credential. Select "Why this label?" on one card in the list of six. Turn on Quiz mode and answer for each card.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
188,801 tokens measured. About $0.53 estimated. The Context tab lists the prompts.
How might we let a site say which AI agents may read it, on what terms, and at what price, so the network can enforce it?
Approach
The site publishes one terms file. The CDN reads it at the edge, checks who the agent is, and answers. The browser can also compare the file with your rules.
Who pays
Publishers who want to get paid for agent reads or to block agents. CDN companies that sell enforcement.
Model
The CDN charges per site, or per paid crawl. Any price is an assumption.
Enforced by
The CDN reads the terms file and the signed agent identity at the edge. It returns allow, 402 payment required, or deny. The agent host must send a verifiable identity.
How to try it
Open the Prototype tab. Pick a sample site, an agent, a CDN, and an identity. Press Play, or press Step to go one move at a time. Read the chat and the "Terms JSON and verdict" panel. Press "Check the terms" to compare the file with your rules.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
194,609 tokens measured. About $0.54 estimated. The Context tab lists the prompts.
How might we make a software update swap that lasts only five minutes impossible to hide?
Approach
Record every update manifest in a public, append-only log. A client accepts an update only if its manifest is in the log.
Who pays
Software vendors that ship a self-update client, and the endpoint security teams that run it. Pricing is per product or per client (assumption).
Model
A vendor pays a monthly fee for each product that uses a hosted log. Security teams pay per client for watching. Both prices are an assumption.
Enforced by
Update clients check that a manifest is in the append-only log before they apply a config. CDNs can mirror the log. Third parties watch it and report a mismatch.
How to try it
Open the Prototype tab. Press "Publish update v1.0.1". Press "Attacker swaps config for 5 minutes". Move the minute slider through minutes 2 to 8. Compare the two clients. Press "Tamper with an old entry". Press "Verify chain" and read which entry broke.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
184,112 tokens measured. About $0.52 estimated. The Context tab lists the prompts.
How might a site tell an AI provider that an agent attacked it, in a way the provider can check?
Approach
A signed abuse report names the agent's signed identity and the evidence. The provider verifies it and can pause the agent in minutes.
Who pays
CDN and bot-management vendors. They sell it as an add-on or per verified report (assumption).
Model
A bundle in enterprise bot management, or a fee per verified report. No price is set here. Any price is an assumption.
Enforced by
The agent host sends the provider token with each request. The CDN checks signatures and sends signed reports. The provider issues the tokens, checks reports, and pauses accounts. The site files the report through its CDN.
How to try it
Open the Prototype tab. Pick a CDN and a provider label. Press Play with the toggle off. Note the time to stop. Turn the toggle on. Press Play again. Check "Change the report on the way" and play once more.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
121,743 tokens measured. About $0.34 estimated. The Context tab lists the prompts.
How might a browser agent know whether an instruction came from the user or from a page?
Approach
The browser labels every prompt with its channel: user, page, extension, or network. It then signs the label. The agent backend checks the seal and refuses page text that claims to be the user.
Who pays
A security lead at a company that rolls out managed browsers with an agent. The vendor is an enterprise browser maker.
Model
A per-seat security tier in a managed browser. Any price is an assumption.
Enforced by
The browser stamps and signs each prompt. The agent backend refuses a sensitive tool call that lacks a signed user tag. Extension stores flag rules that rewrite requests to agent endpoints. This last part is not in the demo.
How to try it
Open the Prototype tab. Keep Seal off. Select the Ask button and read the chat. Select Show the hidden line to see the attack text. Select Reset. Turn Seal on. Ask again. Compare the badges, the notices, and the backend panel. Play the diagram to see where the check happens.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
103,804 tokens measured. About $0.29 estimated. The Context tab lists the prompts.
How might we know that the person who emails, texts, or calls us is a real human, and the one we expect?
Approach
Combine several weak checks into one score. Pass 3 of 5 and you proceed. Never rely on one check.
Who pays
Banks, IT help desks, and managed service providers. They want to stop social-engineering resets.
Model
Per seat, or per verified call. Any price is an assumption.
Enforced by
The messaging app or phone dialer shows the verified badge. The wearable vendor signs the liveness statement. The identity provider holds the passkey.
How to try it
Open the Prototype tab. Pick a surface: email, text thread, or phone call. Pick a scenario. Click Run check on each check, or Run all five. Watch the trust meter and the tally. Read the enforcer panel. See what each party learns. Try the other two scenarios.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
103,218 tokens measured. About $0.29 estimated. The Context tab lists the prompts.
How might we show what an agent will really change before the person approves?
Approach
Compare the request with the agent's plan. Show the difference, like a code diff.
Who pays
enterprise IT and change-management teams. Also agent platform vendors that sell to those teams.
Model
per seat (assumption). The vendor adds plan diffs and receipts to its agent product.
Enforced by
the agent host must show a plan diff before it runs and write an undo record. The repo or ticket system stores the receipt.
How to try it
Open the Prototype tab. Watch the agent answer in the chat. The plan steps appear one by one. Check or uncheck steps. Watch the What will change panel. Click Apply selected, then Undo. Pick the second scenario and repeat.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
183,870 tokens measured. About $0.51 estimated. The Context tab lists the prompts.
How might we help a person understand what they approve when an agent asks?
Approach
Show what will change, in plain words, before the click. Then check whether the person understood.
Who pays
Browser vendors and agent hosts that compete on safety. Also enterprises that require review before an agent acts.
Model
Per seat for enterprise review controls (assumption). Browsers and hosts can also ship it as a free feature.
Enforced by
The browser or host renders the preview from structured tool metadata. It does not use text the agent writes. Sites declare effect summaries in their tool definitions.
How to try it
Open the Prototype tab. Choose a window and an approval style, or keep Random. Press Start a new round. Read the chat. Press Allow or Deny. Answer the three questions and read your score. Repeat with another style. Read the live score and the bars.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
191,740 tokens measured. About $0.54 estimated. The Context tab lists the prompts.
How might five parties agree on what an agent was allowed to buy when something goes wrong?
Approach
The user's mandate, the agent's cart, the merchant's order, and the payment link into one signed bundle. A dispute desk reads it in one view.
Who pays
dispute operations teams at card issuers, acquirers, and payment processors.
Model
a fee per bundle, or basis points on agent-initiated payments, or a seat in dispute software. Every price is an assumption.
Enforced by
the card networks make a bundle a condition of the lighter dispute path. Processors attach the payment block. Merchants sign the order block. Agent hosts sign the cart block.
How to try it
Open the Prototype tab. Press Play, or press Step to go one step at a time. When the dispute opens, read the desk under Without bundle. Press With bundle and read the verdict table. Tick the box that edits the merchant record. Find the block that breaks.
Where to give feedback
Use the Feedback button on the prototype page. Comments are on the admin host after a sign-in with a verified email, when the owner allows it.
Cost to make
120,733 tokens measured. About $0.34 estimated. The Context tab lists the prompts.
People cannot see what an agent did for them, who they are dealing with, or why it acted.
Problem statement
The web has ways to talk, to log in and to move data. It has no common way to answer three questions about AI agents and about the people we deal with: what did I let an agent do, who or what am I dealing with, and why did it act?
Approach
We built 16 small prototypes, each as its own page. Each one shows who would pay and how it gets enforced. Each says when it is a proposal and not a standard.
Hypothesis
We believe each question needs a small, testable proposal, and not one big standard. If it holds, a team can try one idea in minutes and judge it. To confirm: all agents and companies are simulated.
The idea
The web has ways to talk, to log in, and to move data. It has no common way to answer three questions about AI agents and about the people we deal with.
Sovereignty and Permissions: What did I let an agent do, and can I take it back?
Disclosure: Who or what am I dealing with, and where did it come from?
Understandability: Why did it act, and what will change?
Every agent and every company in these demos is simulated. Some ideas are proposals, not standards. Each page says so. Logos are trademarks of their owners. They only name the product, and no company endorses or takes part in these demos.
How each prototype is documented
Each card answers what it is, who pays, how it gets enforced, how to try it, and where to give feedback. Each prototype page has two tabs: Prototype and Context.
The Context tab holds the pain point, the problem, the approach, the hypothesis, the prompts and the cost.
What the build cost
The author runs for all 16 prototypes used 2,585,149 tokens in total, over two rounds. That number is measured. A research run for the second round used 118,066 more tokens. The split into input and output tokens is an estimate, so the cost is an estimate too: about $7.57 at $2 input and $10 output per million tokens for Sonnet 5.5. Planning, review, and publishing by the lead session are not included.
Start your own
The platform has eight starter templates: slides, doc, table, dashboard, notebook, images, feedback, and prototype. The prototype template gives you the two tabs used here. Run fotw new prototype ./my-prototype, then fotw publish ./my-prototype.