Bradley Haven has released Elastifund, an open-source "operating system for agents" that deploys autonomous AI agents to trade prediction markets using real capital. The system, published on GitHub and running on an AWS Lightsail instance in Dublin, uses Python, FastAPI, SQLite, and Elasticsearch as its core stack. Elastic's search and observability infrastructure serves as the <a href="/news/2026-03-14-db9-serverless-postgresql-for-ai-agents">shared substrate for system memory, evaluation, and workflow control</a> across all agent workers — a design choice Haven frames as the central architectural claim: that durable, shared memory and evaluation infrastructure is what separates genuinely self-improving agents from stateless automation. Trading on Polymarket is described as the "first proof lane," with non-trading agent workers like JJ-N (currently offering a Website Growth Audit product) representing the broader platform ambition.
The project's approach to strategy management is where the engineering gets specific. Of 131 strategy hypotheses tested, 12 were formally killed by automated kill rules and two more rejected after hitting evidence-gathering deadlines with zero supporting data. The active BTC5 sleeve shows a historical closed-batch result of +$131.52 across 128 contracts (profit factor 1.49), and a historical calibration validation of 71.2% win rate on 532 resolved markets is cited — though the project explicitly labels this as historical validation rather than a live performance headline. Haven has also published a headline figure of +57.9% return ($247 to $391) across 50 trades, but has been candid in public forums that this came from a single three-hour session on March 11, 2026, with a 51.4% win rate and a profit factor of 1.01, and that the edge has not been replicated. The test suite currently reports 1,961 passing and 50 failing tests, with failures isolated to a single regression file.
The project's regulatory posture carries significant unaddressed risk. Elastifund's primary live trading venue is Polymarket, which paid a $1.4 million CFTC settlement in January 2022 and was required to block U.S. users after its prediction markets were deemed off-exchange commodity options under the Commodity Exchange Act. While the Dublin server location may reflect jurisdictional awareness, it does not insulate a U.S. person from CFTC oversight. The Kalshi integration — a CFTC-designated contract market legally available to U.S. users — introduces a separate compliance layer: operators deploying the system with third-party capital could trigger unregistered Commodity Pool Operator obligations, and the 61-page "Agentic Trading Blueprint" methodology guide Haven sells on Gumroad raises questions about whether the sale constitutes unregistered commodity trading advice under the CEA. The public repo contains no CFTC compliance disclaimers, CTA or CPO registration notices, or legal terms governing downstream operators who fork and run the system with live capital.
The 50-trade sample is too thin to say anything about edge. What the project does offer is an unusually detailed failure record: 12 strategies formally killed by automated rules, two more terminated for failing to produce any supporting evidence before deadline. Most <a href="/news/2026-03-15-vibetrade-claude-trading-agent">open-source AI trading projects</a> don't publish their dead strategies. Haven's longer-term claim is that the same kill-rule architecture and shared evidence layer generalizes to non-trading agents — JJ-N is already running on the platform selling a different product entirely. That's the part of the thesis that hasn't been tested yet.