AgentSmyth

The agents are the product. They read macro data, news tone, factor exposures, options flow and earnings history, test one another's conclusions against the data, and hand back a single view or a trade structure. Arithmetic runs in code rather than in the model, but the routing, synthesis and explanation are model work, and the Harness activates only the agents a question needs. Without them a desk is left with the data feeds it already pays for.

Changes to a portfolio or watchlist wait for the user's approval, and a portfolio analysis returns its proposed change as a pending item rather than applying it. Low confidence outputs go to reviewer queues, and every figure carries its source and calculation. Trade ideas arrive as structured tickets with entry, invalidation and sizing for a person to act on, and the platform describes no automatic execution.

Calculations run in code and agents explain them with citations, so every figure comes with a clickable source and calculation receipt, and lineage can be exported for audit. Morning, midday and end of day briefs are computed ahead of time and render instantly, and the cache resets on corporate actions, macro prints and ratings changes.

Service levels are stated as 99.9 percent availability, under two seconds at the 95th percentile for scheduled briefs and under ten seconds for deep questions. Accuracy figures, back test results for trade ideas and model validation material are not given.

At its 2025 seed round AgentSmyth said 48 institutional clients with $2 billion to $50 billion in assets were using it, and BNY took it into a collaboration program to refine the platform for institutional desks. Its customer stories quote Alex Corner, head of electronic derivatives trading, Alexandra Egorova, a director covering global banks, and Josh Feldman, a senior vice president of engineering, without naming any of their firms.

Corner's desk is said to have cut hours of work to minutes. Feldman recounts an offshore team at a major brokerage accidentally sending 18.4 million requests in one hour, all served with no drops at an average under 5 milliseconds. No client is named alongside a measured result.

Customer data never trains AgentSmyth's models, and no retention and no training flags are passed to model providers. Inference can run inside a customer's VPC or on premises so data stays within the perimeter, and each tenant is isolated by namespace and strict access controls. Detectors catch personal data and secrets at the point of entry, context windows are kept to the minimum, and redaction and policy based blocking sit in front of the models.

The privacy policy, updated May 3, 2026, is issued in the name of AgentSmyth Advisors and covers the websites, applications and dashboards, though its contact details give the company as AgentSmyth Inc. It sets no retention periods, names no subprocessors and says nothing about where data is stored or how it crosses borders. Information a user authorizes when connecting a brokerage or chat platform account is shared under that integration's terms. A data processing agreement covering GDPR and CCPA, residency controls and a subprocessor list are offered only inside a security pack under NDA.

AgentSmyth claims no SOC 2 report or ISO certification. It describes defense in depth, with per tenant isolation, TLS in transit, AES 256 at rest, secrets held in key management or hardware security modules, signed images and supply chain scanning. An annual independent penetration test is run and its summary is shared under NDA, along with a security pack holding a whitepaper, shared responsibility model and control matrix. Role based permissions and admin audit logs govern access.

AgentSmyth's 2025 seed announcement called the company a registered investment adviser and FINRA registered. The SEC's adviser register lists AgentSmyth Advisors, formerly Alta Summit Capital, at 32 Old Slip in New York, the same address as AgentSmyth Inc., with its registration inactive and terminated in March 2026. FINRA's BrokerCheck carries only that adviser record and no broker dealer.

Today AgentSmyth makes no regulatory claim beyond research packaging built with MiFID II and MAR in mind and outputs that carry disclaimers and respect entitlements, while its agents still structure trades with sizing.

Securities and markets are the subjects of this work, so the governance questions are accuracy and which models may answer which questions. Customers choose the models, and AgentSmyth enforces allow and deny rules by use case and region, guards against prompt injection and limits retrieval to allowlisted sources. Outputs carry citations, lineage and risk checks. AgentSmyth gives no evaluation results and no figure for how often outputs are routed to a human reviewer.

Research outputs ship with disclaimers, and AgentSmyth's public terms cover only its website and self serve subscriptions. For a structured trade idea with sizing that a desk acts on, AgentSmyth offers no accuracy commitment, remedy or indemnity.

Customers pick the models and AgentSmyth routes among them under allowlists, but it names no default model or provider. Thomson Reuters Ventures described the platform in 2025 as drawing on more than 100 datasets streaming in real time, and the vendors behind them are not named. The no retention and no training flags go to providers that are not identified either.

AgentSmyth runs as its own workspace or plugs into existing systems through an API, MCP and a command line interface, and a Macro Report API on its developer portal pipes its market view into other platforms. Single sign on with SAML and SCIM and logs sent to the customer's SIEM are supported. Morning briefs are formatted to paste into desk chat, and users can connect brokerage and chat platform accounts. No order management, execution or market data system is named as an integration.

Four deployments are offered. Multi tenant and single tenant SaaS come with regional residency and SIEM export. The customer's own VPC runs over PrivateLink or peering, with the customer's own key management and strict egress allowlists, and on premises installs use Kubernetes, Helm and Terraform from signed images. Hybrid setups keep inference inside the customer's perimeter. The regions offered for the hosted version are not listed.

Pricing starts with a deployment conversation, and no rate is given up front. AgentSmyth's 2024 launch quoted $30 a day for unlimited agent use and $60 a day for agents working on a client's own data, rates it no longer quotes. A May 2025 fulfillment policy describes self serve subscriptions that renew automatically, with refunds considered within 14 days for access problems or a service that differs from its description.

Sales traders, derivatives desks, multi asset teams, portfolio managers and wealth managers are the users. The workflows follow the trading day, from what moved overnight to the pre earnings check and the effect of a rates shock on a book. Coverage spans 5,005 global equities across five or more geographies. The stated clients range from $2 billion to $50 billion in assets, and the customer stories feature bank and derivatives desks. Lending, insurance and banking operations fall outside it.