Kalshi builds AI agent Harrison to review contracts

Kalshi, a CFTC-regulated prediction-market exchange, has developed an internal AI agent called Harrison, built on Anthropic's Claude model, to stress-test contract wording and logic, summarize news, analyze competitors, propose new market listings, and recommend liquidity-provider reward policies. The agent was active as Kalshi cleared nearly $18 billion in notional volume in May and set a weekly record with $5.1 billion during the first week of the 2026 World Cup. New listings still require two human reviewers and a one-to-two hour review window.

Kalshi builds AI agent Harrison to review contracts Prediction-market exchange Kalshi has developed an internal AI agent called Harrison , built on Anthropic's Claude model, to stress-test the wording and logic of prediction-market contracts, according to Bloomberg reporting citing co-founder Luana Lopes Lara. Harrison also handles news summarization, competitor analysis, proposals for new market listings, and recommendations for liquidity-provider reward policies. Benzinga reports the agent was active as Kalshi cleared nearly $18 billion in notional volume in May and set a weekly record with $5.1 billion in the first week of the 2026 World Cup. Reporting frames Harrison as an internal decision-support tool designed to surface contract ambiguities and edge cases across the millions of wagers Kalshi handles daily, with new listings still requiring two human reviewers and a one-to-two hour review window. What happened Bloomberg reports that Kalshi , the CFTC-regulated prediction-market exchange, has developed an internal AI agent known internally as Harrison to help with a range of operational tasks, notably stress-testing the wording and logic of prediction-market contracts. Bloomberg's coverage cites an interview with co-founder Luana Lopes Lara describing the tool's use across the millions of wagers the platform handles every day. Benzinga, citing the same Bloomberg report, notes that Kalshi cleared nearly $18 billion in notional volume in May and set a weekly record with $5.1 billion in the first week of the 2026 World Cup, with Harrison active throughout that period. Several outlets, including Coinness and Valuethemarkets, report that Harrison also performs automated news summarization, competitor trend analysis, proposals for new market listings, and recommendations for liquidity-provider reward policies. Several outlets report the agent is built on top of Anthropic's Claude model. Technical details Per reporting, Harrison is an internal agent layered on a foundation model, described in multiple outlets as Anthropic's Claude. Reported operational capabilities include: - •reviewing and stress-testing contract language to detect ambiguities and edge cases prior to market launch - •summarizing major news items that could affect open markets - •scanning competitor contract designs and market activity for trend signals - •proposing candidate market listings and suggesting liquidity-provider reward policies Bloomberg notes Kalshi maintains over 500 vetted market templates and Harrison suggests which template fits a new event while flagging issues worth a second look. These functions, as reported, are oriented toward automating decision-support workflows rather than serving as a customer-facing product. Industry context Companies operating event-driven markets or high-volume contract systems commonly adopt automation to reduce manual review load and to surface edge cases that human reviewers may miss under scale. For prediction-market platforms specifically, ambiguous contract definitions can create downstream disputes that are costly to resolve and damage platform credibility. Reported deployments of agentic workflows on top of foundation models, as described in the coverage of Harrison, align with a broader pattern where firms use LLMs for contract validation, monitoring, and automated triage. Bloomberg notes that Kalshi previously relied on Yale debate champions to stress-test contract wording - a human approach now supplemented by the AI agent. Context and significance The reported deployment matters to practitioners for three reasons. First, automating contract stress-testing addresses a concrete operational failure mode - contract ambiguity - that is particularly acute for event-based trading. Second, integrating real-time news ingestion with contract checks creates a tighter operational feedback loop between information flows and market definitions. Third, the choice to build an internal agent on an off-the-shelf foundation model reported as Claude illustrates a pragmatic approach many firms use: combine a base model with domain-specific tooling and guardrails rather than training a bespoke model from scratch. What to watch Observers should track the following indicators to assess impact and maturity: - •whether Kalshi publishes technical or governance details about Harrison's scope and failure modes, including how human escalation is handled; - •any regulatory scrutiny or CFTC commentary about automated contract validation workflows on regulated exchanges; - •evidence of reduced dispute rates or faster contract resolution timelines if Kalshi reports operational metrics publicly; - •broader adoption signals from other prediction-market operators or derivatives platforms that reference similar agentic tooling. Implications for practitioners Teams building or operating event-driven marketplaces can view this reported deployment as a case study in applying foundation models to operational risk reduction. Key practitioner considerations include defining clear escalation paths for model outputs, instrumenting contract-edge-case coverage, and measuring false-positive versus false-negative detection rates when validating contract language. Bloomberg notes new listings still require two human reviewers and a one-to-two hour review window. Limitations of the reporting The core facts above are sourced to Bloomberg and corroborating industry outlets. Reporting paraphrases comments from an interview with co-founder Luana Lopes Lara; no verbatim direct quotes were available in the scraped material. The company has not published a technical whitepaper on Harrison. Note: Benzinga discloses a data collaboration agreement with Kalshi. Scoring Rationale A well-documented practitioner case of agentic tooling applied to a concrete operational problem - contract ambiguity on a CFTC-regulated exchange at scale $18B notional volume in May . Relevant to teams building domain agents on foundation models, but not a frontier-model release or industry-shaping event. Practice interview problems based on real data 1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with. Try 250 free problems /problems