Skip to content
Agentic AI

Claude Can Now Run 1,000 Agents On One Job Without Anyone Approving The Launch. TD Bank's Mortgage Agents Show The Agentic Workflow Automation Pattern That Works In A Regulated Back Office

On 9 October 2026 Anthropic put dynamic workflows into beta for Claude Managed Agents: one agent writes a program that fans a job out across up to 1,000 agents, and the server runs it with no step at which a developer approves the launch. On 13 October 2026 TD Bank said it had expanded agentic AI across mortgage funding and discharge, with every output reviewed by a colleague and an earlier agent cutting average processing time from 15 hours to under three minutes, by the bank's account. This article puts the two together and sets out the template for a prop firm's month-end review, a broker's re-papering or a fintech's KYC refresh: agents read and cross-check at scale, rules stay in code, a named person signs.

AlchmAI Editorial8 min read

1,000

Agents one Claude workflow run may start, 64 threads at once, 24-hour default lifetime (Anthropic docs, 9 October 2026)

66 of 70

Planted bugs a dynamic workflow found in a 116,000-line codebase, against 14 to 27 for a single agent (The Decoder, 9 October 2026)

15 hrs to <3 min

TD Bank mortgage pre-adjudication processing time after its May 2026 agent, a company claim (TD, 13 October 2026)

52%

Financial firms already experimenting with agentic AI (Cambridge Centre for Alternative Finance, 28 April 2026)

Agentic workflow automation got its biggest capability jump and its best governance example in the same week. On 9 October 2026 Anthropic added dynamic workflows to Claude Managed Agents, in beta: an agent can now write a program that fans one job out across up to 1,000 agents, and Anthropic's servers execute it as a workflow run, with no step at which a developer approves the launch. On 13 October 2026 TD Bank announced it had expanded agentic AI across its mortgage funding and discharge operations, with every output going to a colleague for review. Together they show the pattern that works in a regulated back office: agents read, cross-check and draft at scale, the rules stay in code, and a named person signs anything that matters.

The work TD describes is the work a prop firm, a broker or a fintech does every month. Reviewing hundreds of funded-trader accounts against a rulebook, re-papering client agreements, refreshing know-your-customer (KYC) files: a pile of documents, deterministic checks and a decision somebody has to own. The number to watch is therefore not 1,000 but the budget field, the 24-hour lifetime and the fact that stopping a run means asking the agent to stop it.

What Did Anthropic Actually Ship On 9 October 2026?

Dynamic workflows arrive under the managed-agents-2026-04-01 beta header, for "work with many pieces, such as reviewing hundreds of documents". The orchestration guide lists the intended uses as "audits, migrations, deep research, and cross-checking". The limits page holds the operational facts. A run may have 64 threads working at once, and the documentation adds: "The API doesn't guarantee this number." A workflow may start 1,000 agents over the run's whole life. A run lives 24 hours by default, or whatever lifetime the agent itself sets, and "No event says what lifetime the agent set."

The Decoder, in a 9 October 2026 article by Matthias Bastian, reports a test in which 70 bugs were hidden in a 116,000-line codebase: "a single agent caught between 14 and 27 per run" while "the dynamic workflow consistently hit 66". The Decoder attributes the test to Anthropic, which has not published it as far as we can find, so it is a reported result, not a benchmark. The Decoder also writes that "Up to 1,000 agents can run in parallel per execution", which the limits table does not support: 1,000 is the lifetime count, 64 the parallel count.

Who Approves A 1,000-Agent Run?

Nobody, by design. Anthropic's orchestration guide states it in a parenthesis: "Permission policies apply to the tools that a run's agents call, not to starting the run." MIXED, in an 11 October 2026 analysis by Shane S. Ellison, put it more bluntly: "Nothing in the flow asks a developer to approve the launch." The decision to spin up 800 agents on a document set is the agent's, steered only by its system prompt.

The controls that exist are financial and temporal, not procedural. A session budget caps spend at list rates. When the cap is reached every open run pauses, although each thread finishes the request it already started, so "a run can pass the budget by one request for each working thread". An interrupt "ends no run"; to stop a run you send a message asking the agent to stop it, and that message is rejected while the session is paused at its budget. A paused run's lifetime keeps passing, so a run left paused for a day ends with a timeout_error.

How Does TD Bank Use Agentic Workflow Automation In Mortgages?

TD's 13 October 2026 release, datelined Toronto, describes capabilities in mortgage funding and discharge that "coordinate multi-step document verification, connect information across systems, and generate outputs for colleague review", "designed to double colleague productivity in select mortgage processes", with the initial funding review done "in up to half the time". The sentence that matters most is this: "Colleagues review AI-generated outputs and remain responsible for decisions requiring human expertise and judgment." All figures are company claims.

The foundation is the pre-adjudication agent TD launched in May 2026, which the 13 October release says "reduced average processing time from 15 hours to less than three minutes". A TD Stories article of 31 July 2026 explains what it does: it scans a client's documents, calculates income, checks it against policy requirements, looks for discrepancies and generates a memo for underwriters. The same article says: "Currently, 100% of new personal mortgages go through this process." The agent reads and reconciles; policy is a check the agent runs against, not a judgement it makes; the output is a memo to an underwriter, who decides.

Customer-facing banking follows the same rule. Backbase's Conversational Banking, generally available from 8 October 2026, lets customers ask a bank's agents to act for them, with one condition in the release: "Any consequential request requires explicit customer confirmation." Backbase says one of the four largest South African banks raised self-service containment from 20% to 70% on it. Vendor figures, but the same principle: the agent does the work, a human confirms the consequence.

Our View: Count The Signatures, Not The Agents

The Cambridge Centre for Alternative Finance's 2026 Global AI in Financial Services report, published on 28 April 2026 from 628 organisations, found 52% already experimenting with agentic AI, while only 40% reported higher profitability from AI and 43% reported no change. Running a thousand agents is now a commodity. What turns it into a result is the structure around it, and TD's structure has three layers a firm of 20 people can copy.

  1. 01Agents do the reading, cross-checking and drafting. This is where fan-out pays, and where a wrong answer is caught downstream rather than acted on.
  2. 02Deterministic rules stay in code. Whether a trader breached a daily loss limit, whether a KYC file is past its refresh date, whether an income figure meets policy: these are calculations and lookups. A rule that lives in a prompt is a rule nobody can audit.
  3. 03A named person signs anything consequential. A payout refused, an account closed, a mortgage funded. The agent's memo cites its sources; the person's name goes on the decision; the log keeps both.

Anthropic's release is powerful and incomplete in the same breath: powerful because a prop firm with 2,000 funded accounts to review at month-end should not wait for one agent to read them in sequence, incomplete because the launch has no gate and the stop is a request to the agent. Firms must add the maker-checker step TD has and the platform does not: the run's output is a proposal, a human's signature is the action. Budgets and a tested stop path matter as much as the agent count, because the question a regulator asks after an incident is not how many agents ran but who could have stopped them, when, and for how much.

“A thousand agents can read the files. The rulebook still has to live in code, and a person still has to sign. The firms that get agentic workflow automation right decide those two things before they decide the agent count.”

What It Means For Trading And Fintech Firms

  • Prop trading firms: structure the month-end review of funded-trader accounts like TD's pre-adjudication. A rule engine decides every breach from tick data; agents read trade history and draft a memo per flagged account; a named risk manager signs every refused payout. Set a dollar budget per run and record who may raise it.
  • Brokers and CFD providers: use fan-out for agreement re-papering after a terms change. Agents extract the clauses that differ from the new template and draft the letter; code checks each file against the checklist; the compliance lead signs each batch. Require human confirmation on any tool that writes to the CRM or sends mail, because that is where permission policies apply.
  • Fintech and trading startups: make a KYC refresh the first agent workflow you build. Agents gather and cross-check the evidence, rules decide whether a file passes, and the money laundering reporting officer signs the exceptions. Tie the budget to the case count.
  • Research and quant teams: dynamic workflows suit data audits and cross-checking many sources. Set the run's lifetime in your own code, and test the stop path before the run touches a production dataset.
  • Compliance leads: ask three questions. Where is the rulebook, in code or in a prompt? Whose name is on the decision? What stops the run, and has anyone tested it?

The Bottom Line

Anthropic's dynamic workflows, in beta from 9 October 2026, let one Claude agent start up to 1,000 agents on a single job, with a dollar budget that pauses the run at budget_reached and no step at which a developer approves the launch. TD Bank's 13 October 2026 expansion of agentic AI across mortgage funding and discharge shows the other half: colleagues who review every output and own every decision. For a prop firm's month-end review, a broker's re-papering or a fintech's KYC refresh, that is the template for agentic workflow automation: fan the reading out, keep the rules in code, put a name on the signature, and set the budget and the stop path before the agent count. It is the work we do at AlchmAI, a London trading technology and AI studio, with approvals, limits and an audit log designed in before the first run.

References & Further Reading

Agentic AIagentic workflow automationTrading Workflow automationWorkflow Automation LondonAI AutomationClient-Facing & Internal ToolsClaude Managed AgentsTD Bank
Share Email
AI

AlchmAI Editorial

Research and analysis, London

The AlchmAI team writes about the markets, technology and regulation we work with every day. We build trading platforms, real-time charts and AI analysis tools for brokers, prop firms and fintech teams from our office in Mayfair, London. Every article lists its sources. Nothing we publish is investment advice.

This article is general information and commentary. It is not investment advice or a recommendation to buy or sell any investment. Important information

Keep reading

All articles