AI Agents

Schwab Blocks Agents Weekly: Build for Counterparty Denial

Rushil ShahRushil Shah
14 min read
Share

Schwab reports blocking agent attempts on client money every week and JPMorgan bars outside agents entirely, while Adobe opens Creative Cloud via MCP. Denial is now a design input, not a news story.

TL;DR

Counterparty denial is now a permanent property of the web, not a temporary hiccup to route around. Map every workflow step to one of five access tiers — sanctioned API/MCP, licensed data, delegated authorization, signed or unsigned browser access, human-in-the-loop — and build the denial path before the happy path. For money movement and licensed data, the correct 2026 build is a human-gated handoff with a state-preserving resume, not an autonomous agent. Anything else is a liability transfer onto your user.

What actually changed in the last month

Three days in mid-October reset the assumption that an agent with a user's password is a legitimate user. On 13 October, CEO Jamie Dimon said JPMorgan Chase will not allow outside AI agents to access customer accounts and operate on their behalf, citing risk to the firm and its customers. Three days later Charles Schwab CEO Rick Wurster told Bloomberg TV that the broker stops attempts by major AI agents to reach client money every week and that he receives a weekly report on them, described as being in the tens each week. Treat that figure as what it is: a number given verbally in a TV interview and repeated in trade press, not a disclosure in a filing. It is directionally useful and nothing more.

The pattern is not banking-specific. Reporting on 17 October catalogued agents hitting walls at Delta and Yelp as well as the banks, with Yelp admitting only agents enrolled in its paid data-licensing programme and Amazon blocking Meta's Muse since 20 September. Citi and Bank of America moved in the same window, and JPMorgan's CFO framed agents partly as a deposit-retention and customer-experience problem, which tells you the motive is not purely security.

And in the same month, the opposite move: at MAX, Adobe said it will open Photoshop, Illustrator, After Effects and Premiere to third-party assistants such as ChatGPT and Claude, using local MCP integrations. So the landscape is not closing. It is splitting. Counterparties whose product is the workflow are opening sanctioned lanes; counterparties whose product is the data or the money are closing credential lanes and selling or withholding access instead.

Weekly Schwab report of blocked agent attempts to reach client money Source: Notebookcheck summarising Bloomberg TV, Oct 2026
Oct 13 JPMorgan states outside agents cannot operate customer accounts Source: CNBC, 2026
Jul 1, 2026 Cloudflare starts labelling verified bots Direct vs Intermediary Source: Cloudflare Docs, 2026
11 bot behaviour categories in Cloudflare's taxonomy, including a separate “Transact” class Source: Cloudflare Docs, 2026

The counterparty access tier model

Stop asking “can our agent do this?” and start asking “under which tier is this step permitted, by whom, and what happens the day that tier is revoked?” Every step in a workflow sits in exactly one tier, and the tier is a property of the counterparty, not of your stack. A single procure-to-pay flow routinely spans four of them: sanctioned API at the ERP, licensed data at the price-check step, browser access at a small supplier with no API, and human-gated at the payment release.

TierHow access is grantedRealistic time to obtainPrimary failure modeSafe scope today
1. Sanctioned API / MCP serverPublished docs, developer agreement, OAuth app reviewDays to weeksScope creep: the API exposes 60% of what the UI doesRead and write within documented scopes
2. Licensed dataCommercial licence, per-call or per-seat pricingWeeks to months, legal involvedCost scales with agent retries, not with valueRetrieval, enrichment, comparison
3. Delegated authorizationCounterparty-sanctioned token tied to a named third party, scoped and revocableMonths, and often unavailable to youRead scopes granted, action scopes withheldRead-only account and order data
4a. Signed browser accessCryptographic agent identity (Web Bot Auth), site chooses to allowWeeks, plus per-site allowlistingVerified and allowed are not the same thingPublic-surface navigation and reads
4b. Unsigned browser access with user credentialsNobody grants it; you take itImmediateTerms breach, liability shift to the user, silent blocksInternal systems you own, and little else
5. Human-in-the-loop handoffYour own product designImmediateHandoffs with no context packet get ignoredMoney movement, irreversible or regulated actions

The useful discipline is to write the tier next to every step in the workflow spec and refuse to let an unlabelled step ship. When we scope agent integrations, the access register is the first artifact produced, before any model choice. Model capability is almost never the binding constraint — see model coverage if you want the capability side — counterparty permission is.

Design the denial path before the happy path

The dangerous denial is not HTTP 403. A 403 is a gift: it is explicit, loggable and easy to branch on. The expensive ones are the quiet failures. A 200 response carrying a consent interstitial. A search page that returns results for a human session and an empty array for yours. A checkout that accepts the first POST, then step-ups to SMS verification the agent cannot see. A rate limiter that does not reject but degrades, returning stale cached data your agent happily reasons over. Browser agents are especially prone to this because the DOM renders, the screenshot looks plausible, and the model confabulates success.

So instrument denial as a first-class outcome with a typed taxonomy, not a generic exception. At minimum: hard_block, rate_limited, challenge_required, step_up_auth, content_degraded, tier_revoked. Each type gets its own policy. Only rate_limited is allowed to retry, and never on a side-effectful step.

Runtime behaviour on a denied step

1
Classify, do not retry

Map the response to a denial type using status, headers, content fingerprint and a cheap assertion on expected payload shape. A result that passes schema validation but fails a freshness or row-count assertion is content_degraded, not success.

↓
2
Checkpoint state

Persist the run as a resumable object: inputs, prior step outputs, the idempotency key already issued, the exact request that was denied, and the evidence bundle (status line, response headers, screenshot hash). State must survive process death, because the human will come back hours later.

↓
3
Degrade one tier, once

Try the next permitted tier exactly once. If the licensed feed is down, do not silently fall back to scraping the same vendor — that converts an outage into a contract breach. Fall back only along paths the access register declares legal.

↓
4
Hand off with a context packet

The human gets the goal, what already happened, what is blocked, the exact action required, and a one-click resume link. A Slack message saying “run 4417 failed” is not a handoff, it is a notification that will be ignored.

↓
5
Resume idempotently

Resume must re-check whether the blocked action was partially completed before re-attempting it. Reuse the original idempotency key. Assume the counterparty may have accepted the first attempt and blocked only the confirmation read.

!

Treating a block as a transient error and retrying with backoff

Generic HTTP clients and most agent frameworks ship with retry-on-failure defaults. Against a bot-management edge, a retry storm is the exact signature of an attack: you escalate from a soft challenge to an IP-range ban, and you may burn the shared egress IPs of your browser-automation vendor for every other tenant. On side-effectful steps you also risk duplicate submissions, because the denial often lands on the confirmation read rather than the write.

Fix: disable framework-level retries on any step that mutates counterparty state, retry only on an explicit 429 with a Retry-After value you honour, cap total attempts per counterparty per hour, and route everything else to the human path with state preserved.

Where the line sits in 2026: money movement stays human-gated

Here is the part most vendors will not say plainly. For a defined class of steps, the right build today is not an autonomous agent. The class is: anything that moves money out of an account, anything that creates a binding commitment under someone else's identity, anything whose counterparty terms prohibit third-party credential use, and anything over licensed data where your access depends on a contract you have not signed.

That is not timidity, it is risk allocation. Reporting on the October bank moves notes that where a customer hands an agent their credentials, the liability is pushed back onto the customer. If you ship a product that encourages that, you are building on a foundation whose failure cost lands on your user and whose reputational cost lands on you.

The line per layer, in practice:

  • Read of account state: automate, via tier 1–3 only. A read is recoverable; a wrong read is a bad answer, not a wire transfer.
  • Draft and stage an action: automate fully. Building the payment batch, the trade ticket, the PO amendment is where the hours actually are.
  • Authorise and execute: human, with a signed approval record. The agent prepares, the person presses.
  • Reconcile and file: automate, with assertion-based verification rather than model self-report.

Roughly, the preparation steps carry most of the duration and the authorisation step carries most of the risk. Automating the first and gating the second buys you the majority of the cycle-time win with a risk profile you can put in front of a compliance reviewer. That is the shape of nearly every durable build we document in case studies.

!

Putting user bank credentials in your agent's secret store

It works in the pilot, which is the problem. You now hold reusable credentials for a counterparty whose terms forbid third-party login, with no revocation path that does not involve the user changing their password, no audit trail the counterparty will recognise, and a detection surface that gets better every quarter. The first incident is not a technical failure, it is a contractual one.

Fix: never store a counterparty credential your contract does not explicitly cover. Use sanctioned delegated tokens where they exist; where they do not, keep the credential in the human's own session and have the agent prepare the action for the human to execute inside that session.

Getting sanctioned: identity, licence, contract

There is now a real mechanism for being a recognised agent rather than an anonymous one. Cloudflare defines a verified bot as one that identifies itself honestly — via a cryptographic Web Bot Auth signature, a published IP list with a stable user agent, or reverse DNS — and does not abuse the access that honesty earns. Implementation is concrete: generate an Ed25519 key, publish a JWKS at a fixed well-known path, sign requests with RFC 9421 HTTP message signatures, and register the directory.

bash
# Ed25519 keypair for Web Bot Auth request signing
openssl genpkey -algorithm ed25519 -out private-key.pem
openssl pkey -in private-key.pem -pubout -out public-key.pem

# Publish the JWKS here, over HTTPS, signed, with
# Content-Type: application/http-message-signatures-directory+json
#   /.well-known/http-message-signatures-directory

Two operational details the quickstarts underplay. First, Cloudflare does not support every RFC 9421 component: signing @query-param or including parameters such as sf, bs, key or req will fail verification outright, and an expires value that is too tight will have lapsed by the time the request reaches the edge — roughly a minute is a sane floor. Second, and more strategically: verification is identity, not permission.

Cloudflare now labels verified agents Direct or Intermediary, where an intermediary is an agentic service that many different end users can operate, creating transitive trust — the site may trust the operator but not every end user behind it, with experimentation around forwarding end-user information using the RFC 7239 Forwarded header. If you build on a hosted browser or a general-purpose agent platform, you inherit that platform's reputation and its noisiest tenant's behaviour. Running your own signed identity for high-value counterparties is worth the operational overhead precisely because the reputation becomes yours.

On the commercial side, the Yelp posture — access conditional on a paid data licence — is the template. When you negotiate one, the clauses that matter operationally are rarely the ones legal starts with: a written rate limit and burst allowance, a named user agent plus key-rotation procedure, advance notice before access-policy changes, an explicit statement of whether degraded responses may be served instead of errors, and the right to export your own call logs for audit. Ask for the denial semantics in writing. “What does your edge return when we exceed quota” is a question most vendors have never been asked and most procurement teams now should.

Denial telemetry and the evidence procurement now asks for

Once agents touch external counterparties, buyers stop asking about accuracy and start asking about provenance and control. Four artifacts answer nearly every question in a security review.

The access register. One row per counterparty per step: tier, legal basis (contract, licence, published API terms, user's own session), credential owner, revocation path, and the date the terms were last read. This is a living document, not a one-off; the October bank changes invalidated rows in a lot of registers overnight.

Denial rate by counterparty and type. Track hard_block, challenge_required and content_degraded separately and alert on trend, not threshold. A counterparty's posture usually shifts gradually — more challenges, then soft degradation, then hard blocks — and the gradient is your early warning that a tier is about to be revoked.

Human-gate metrics. Time from handoff to decision, percentage of handoffs abandoned, and resume success rate. A workflow with a 40% abandonment rate at the gate is not a working workflow, it is a queue.

Per-action evidence bundles. For every consequential action: who or what initiated it, which tier was used, the authorisation record, the request/response pair, and the approving human's identity. This is the single thing that most shortens an enterprise security review, and it is far cheaper to emit from day one than to retrofit.

The decision rule for this month

The rule

Ship the step autonomously only if (a) the counterparty has a sanctioned tier you are actually enrolled in, (b) the action is reversible or idempotent, and (c) a denial leaves the system in a state a human can resume without re-doing work. If any of the three fails, build the handoff instead and keep the agent on preparation. Revisit per counterparty each quarter, because the tiers are moving in both directions.

Apply that rule and the October news stops being a setback. Adobe's MCP opening and the banks' credential closure are the same signal read from two sides: access is becoming explicit, named and contractual. Teams that built on undeclared browser automation are rewriting. Teams that built the tier model and the denial path are mostly changing table rows. If you are mid-pilot and discovering blocks, the work this month is not a better scraper — it is the access register, the denial taxonomy and the resume path. We help teams do exactly that scoping; start here if you want a second pair of eyes on which of your steps should not be autonomous yet.

Frequently Asked Questions

Why are banks blocking AI agents if the customer consented?

Because consent to the agent is not consent under the bank's terms. JPMorgan said in October it will not allow outside agents to access accounts and act on customers' behalf, citing risk to the firm and customers. Account agreements generally require the customer to access the account directly, so a third party logging in with shared credentials breaches the agreement regardless of intent, and the bank cannot distinguish that session from credential theft.

Will signing my agent with Web Bot Auth stop the blocks?

It stops anonymous blocks, not policy blocks. Signing proves identity using an Ed25519 key, a published key directory and RFC 9421 message signatures, which moves you from unidentified traffic into a recognised category. The site still decides whether that category is allowed. Cloudflare additionally distinguishes agents operated by a single party from intermediary services driven by many end users, and a counterparty can allow one and refuse the other.

What is the single most common architecture mistake in blocked-agent pilots?

Treating denial as an exception rather than an expected outcome. Teams wrap the call in a try/catch with retries, lose all intermediate state when it finally fails, and hand the human a stack trace. The fix is a typed denial taxonomy, a checkpointed run object, a single controlled tier fallback, and a handoff that includes goal, progress, blocker and a resume link.

Is it ever acceptable to use a customer's credentials in a browser agent?

For systems your organisation owns and controls, yes, with normal secret management. For third-party financial or regulated platforms, no. Terms typically forbid it, detection is improving, and reporting on the October bank moves indicates liability is pushed onto the customer who shared the credentials. If a sanctioned delegated tier does not exist, keep the credential in the human's own session and have the agent prepare the action rather than execute it.

How do I detect a silent block rather than a hard 403?

Add assertions that a degraded response would fail: expected minimum row count, a freshness timestamp, presence of a known control record, or a checksum against a second source. Fingerprint response bodies to catch consent walls returned with a 200 status. Never let the model's own report of success close a step; the verification must be deterministic code outside the model.

Does Adobe opening Creative Cloud to third-party agents signal the blocks will reverse?

It signals divergence, not reversal. Adobe is opening Photoshop, Illustrator, After Effects and Premiere to assistants such as ChatGPT and Claude through MCP, because agent access makes its software more valuable. Counterparties whose asset is the data or the money have the opposite incentive and are gating or charging for access. Expect the sanctioned lane to widen for tool-like products and stay narrow for custodial ones.

What should I ask a data provider before signing an agent access licence?

Four things beyond price: the written rate limit and burst behaviour; exactly what the edge returns when you exceed it, including whether degraded or cached data may be served instead of an error; notice period before access-policy changes; and your right to export call logs for audit. Also confirm whether the licence covers downstream caching and how long you may retain retrieved data.

AI agentsagent accessworkflow automationagentic commercebot managementWeb Bot Authfintechhuman-in-the-loopMCPenterprise AI

Published

AI-assisted writing · Reviewed by the Twarx research team

Share:
Share

Research digest

AI Research Briefing

Honest insights on AI agents, Small Language Models, and local RAG. No hype. Only when we have something worth sending.

  • No hype, just measurable outcomes
  • Read by 2,400+ engineers
  • Unsubscribe anytime

Continue reading

More from Articles

AI Agents

AI Agents for Ecommerce Operations: The 2026 Readiness Guide

16 min read
AI Agents

n8n or a Custom Agent Stack for Ecommerce Operations?

16 min read
AI Agents

One Agent Won't Fix Marketing Ops. Five Layers Might.

16 min read
AI Agents

AI Agent for Finance Operations Automation: The 2026 Production Playbook

14 min read