AgentLeash

The control plane for your autonomous AI shopping agents — set budget caps, approval thresholds, and blocked categories once, then watch every agent purchase get approved, held, or blocked against your policy in real time.

Date: 2026-05-31 Form factor: Web app (single-page; mobile-friendly) Status: Prototype

What it is

AgentLeash sits between your AI shopping agents and your card as a policy gate. You connect the agents already buying on your behalf in 2026 — ChatGPT Instant Checkout, Perplexity Comet, Amazon Rufus, Google's UCP checkout — set a handful of household rules, and AgentLeash evaluates every purchase attempt the instant it happens: approve it, hold it for your one-tap sign-off, or block it outright. The result is a single dashboard where you can finally see, cap, and audit what your agents are spending.

The prototype demonstrates the end-to-end flow on one household's day: 17 real-world agent purchase attempts streaming in across four agents, scored live by an in-browser policy engine you can re-tune with sliders and watch every decision recompute.

Who it serves

Households and individuals who have started letting AI agents buy things for them — a fast-growing group in 2026, when 45% of consumers already use AI for part of the buying journey and 70% are at least somewhat comfortable with an agent purchasing on their behalf (IBM IBV, NRF 2026). The specific pain: handing an agent a real card number means the number lives in the agent's config with no built-in spending cap — if it makes a mistake, gets prompt-injected, or enters a processing loop, there is no automatic safeguard. Concrete personas:

• The busy parent who let ChatGPT and Amazon Rufus handle grocery and household refills and now has no single place to see the total or stop a runaway reorder.
• The early adopter running three or four agents at once who wants electronics and travel purchases over $75 to pause for a human glance, while $40 grocery baskets just clear.
• The security-conscious user who has read about agent prompt-injection and wants cash-equivalent buys (gift cards), gambling, and brand-new subscriptions blocked or held by default.

Why it could be profitable

Monetization is freemium consumer SaaS with a family tier and a B2B card-issuer API:

• Free: One connected agent, a monthly cap, and the decision feed. No account required for the demo.
• Pro ($8/mo): Unlimited agents, per-category and per-merchant rules, the approval queue with push/SMS alerts, velocity limits, new-subscription guard, and audit-log export.
• Family ($14/mo): Per-member sub-budgets, kid/teen agent scopes, shared approval inbox.
• B2B issuer API ($0.02–0.05 per evaluated transaction, or rev-share): Banks, neobanks, and card issuers embed AgentLeash as the "agent spend control" layer their customers will demand the moment agentic checkout is default. This is the larger pool — every issuer needs a guardrail story for delegated agent payments, and most have none.

The timing is the whole thesis. Google announced the Universal Commerce Protocol at NRF on January 11, 2026, standardizing AI-completed checkout across merchants; Juniper Research (April 2026) projects ~$8B in agentic commerce spend in 2026, scaling toward $1.5T by 2030. The agents shipped first; the guardrails did not. Privacy.com and Oracle both flagged in 2026 that agents handed a raw card have no spend ceiling and no automatic stop. AgentLeash is the missing control plane for the exact moment delegated spending goes mainstream — and consumer trust is the bottleneck (only 24% trust AI recommendations outright), which is precisely what a visible, enforceable guardrail unlocks.

Form factor & scope

Single-page web app, sized for mobile and desktop. Scope-locked to the policy + audit layer — AgentLeash does not itself broker payments, issue cards, or run the agents; it is the decision gate and ledger that sits in front of them. The minimum viable scope demonstrated here:

1. See your connected agents; pause any one (or freeze all) with a toggle.
2. Set the policy — monthly cap, ask-first threshold, per-hour velocity limit, blocked categories, blocked merchants, new-subscription guard.
3. Watch the live decision feed score all 17 purchase attempts in chronological order, each tagged Approved / Held / Blocked with the exact rule that fired.
4. Work the pending-approval queue — approve or deny held items and watch the budget, KPIs, and feed update.
5. See spend-by-agent and export a plain-text audit log.

How to run it

1. Open index.html in any modern browser (Chrome, Firefox, Edge, Safari).
2. Drag the Monthly cap, Ask-first, and Velocity sliders, toggle the blocked-category chips, pause an agent, or hit Freeze all — every decision in the feed and queue re-evaluates instantly.
3. In the Pending approval queue, click Approve or Deny on a held purchase and watch the approved-spend total and KPIs move.
4. Use Copy audit log / Download .txt to export the full policy + decision log.

No build step, no API keys, no accounts. Sample data is embedded inside index.html as a <script type="application/json"> block so the page works directly from file:// with no local server. A standalone copy of the same data also lives at sample-data.json in this folder.

What's in this prototype

• A live policy engine (script.js) that evaluates each purchase in chronological order against: emergency freeze, paused-agent, blocked-merchant, blocked-category, per-agent velocity limit, new-subscription guard, ask-first threshold, and running monthly-cap — first rule to fire wins, with a human-readable reason.
• Four modeled agents — ChatGPT Instant Checkout, Perplexity Comet, Amazon Rufus, and Google AI Mode (UCP, shipped paused) — each with its own scope note and a pause toggle.
• 17 purchase attempts crafted to exercise every rule: routine groceries that clear, a $329 headphone buy held at the threshold, a DraftKings top-up blocked on category, a GiftCardVault buy blocked on merchant (the classic agent-exploit pattern), a fourth consumable reorder in 30 minutes tripping the velocity limit, and a brand-new Spotify subscription held by the subscription guard — netting 10 approved, 4 held, 2 blocked at the default policy.
• Editable policy controls — sliders for cap / threshold / velocity, toggle chips for blocked categories, and switches for the subscription guard and emergency freeze, all wired to instant re-evaluation.
• A pending-approval queue with working Approve / Deny overrides that feed back into the budget math.
• KPIs + budget bar (approved MTD, remaining, held, blocked-and-prevented) and a spend-by-agent breakdown.
• A plain-text audit export (copy or download) capturing the active policy and every decision — the artifact you'd hand a bank or keep for your own records.

Roadmap

• Real agent connections via the Universal Commerce Protocol and the major agent SDKs, replacing the sample feed with live purchase webhooks.
• Push / SMS approval so a held purchase pings your phone and clears with one tap.
• Per-member and per-agent sub-budgets, plus kid/teen scoped agents for the Family tier.
• Anomaly detection beyond static rules — flag spend that deviates from each agent's learned pattern (merchant, time-of-day, amount).
• A B2B issuer SDK that drops the same policy engine into a bank app as the "agent spend control" panel, billed per evaluated transaction.
• Prompt-injection-aware blocking that cross-references known agent-exploit merchant and category patterns (cash-equivalents, freshly registered sellers).

Sources

• Google Cloud — A new era of agentic commerce — Universal Commerce Protocol (UCP) announced at NRF, January 11, 2026.
• McKinsey QuantumBlack — The agentic commerce opportunity — $3–5T in agent-orchestrated retail spend projected by 2030.
• Ekamoira — What Is Agentic Commerce? 45% of Shoppers Use AI (2026) — IBM IBV NRF 2026 figures: 45% use AI in the buying journey, 70% comfortable with agent purchases, 24% trust AI recommendations outright.
• Privacy.com — AI Agent Payment Solutions in 2026, Compared — handing an agent a raw card number stores it in config with no built-in spending controls.
• Oracle — Runtime Budget Guardrails for Agentic AI — the case for spend thresholds, scoped permissions, and post-purchase controls on autonomous agents.
• AgentPMT — The $15 Trillion War for AI Shopping Agents — guardrail infrastructure (spend caps, vendor whitelisting, real-time monitoring) framed as what makes autonomous agent commerce viable.