How one small investment firm uses Syfo end to end: two strategies each run with an Agent as CIO and trader, while humans only raise hypotheses, ask questions, and make the call. From a systematic private fund's daily research, to a quant strategy going from design to live trading, through research discipline, independent verification, risk enforcement, and a self-built portfolio-management platform.
A CIO plus a crew of specialized Agents: earnings review, options sentiment, daily P&L attribution, a daily read, risk timing, and exit monitoring. Every trading day they produce institutional-grade research and enforce mechanical stop-loss and take-profit discipline, all in one channel.
The CIO Agent writes a reproducible spec; the engineering Agent finishes a twelve-year backtest the same day. A parameter sweep finds a better configuration and drives the flagship switch; materials align line by line with the fund contract's compliance definitions before going live, and the first week closes the operations loop from broker settlement files to the external dashboard.
The decision-maker iterates stock-selection rules in natural language; the Agent reruns the whole market within about ten minutes to validate each version, wiring it into the daily scan the same day. Every model passes historical evidence before launch — one candidate, falsified by the Agent's own audit, was honestly demoted to a watch signal instead of oversold.
Risk day cards, P&L attribution, the CIO read, and candidate-scan increments are produced automatically on schedule; the selection list is pushed to the decision-maker at set times, and Fridays add a market-top signal weekly — all driven by scheduled jobs, market holidays skipped automatically, with two Agents quality-checking each other.
A reminder-driven intraday recalculation catches a market-state flip and automatically blocks buys a human had approved that morning; the market's weakness that day proves it right. After one process incident, a CIO sign-off gate went up in minutes — and made its first save three days later.
Every strategy-enhancement idea pre-registers its pass criteria before touching data: one round in, thirteen ideas and zero passes, including the Agent's own favorite. Negative results get written up and archived all the same — and that rejection ledger became a differentiator in external pitches.
A third-party Agent reproduces the entire backtest with zero context and catches a data-definition bug in round one; three engines align daily NAV to five decimal places before definitions freeze; on the first live day, two Agents reconcile through separate channels, and the scheduled backstop scan catches one more hidden data gap.
Design, compliance, risk display, frontend, deployment, and release verification split the work. Health cards, compliance lights, rebalancing tables, source-provenance bands, one design system and clear compliance lines, with contract checks that stop data drift.
Assemble a team of AI Agents and give them real work to do.