Video summary

How I manage 250+ AI Agents for my Hedge Fund

Main summary

Key takeaways

Technology

Summary

The presenter describes running roughly 250 concurrent AI-agent sessions for an agentic hedge fund and related work. He says the system has no employees, consultants, or agencies. His role is mainly to set priorities, make product decisions, and review selected changes.

Agent roles, skills, and infrastructure

  • Define the problem before recruiting agents. A custom “agent recruitment” workflow interviews him about a task, breaks it into roles, and produces agent skills. He treats each skill as a standard operating procedure (SOP): clear instructions for a repeatable job.
  • Run work across three machines:
    • A MacBook Pro for hands-on work, typically managing 10–20 chats.
    • A Mac mini running about 150 scheduled automations, from hourly to weekly.
    • An M2 Max running around 100 research agents that run and analyze backtests locally.
  • In his research workflow, agents propose test configurations, run backtests through an API, and evaluate results against measurable criteria such as drawdown and portfolio return.
  • He prefers a main orchestrator agent that delegates to subagents, keeping its own context relatively uncluttered. He compares the workflow to a Zapier, Make, or n8n automation, where each step can be handled by a skill or subagent.

Shared memory and autonomous development

Agents share a simple JSON ledger that records prior work. Scheduled agents consult it to avoid repeating tasks or to resume where earlier agents stopped.

For software maintenance, agents can inspect a repository, find and fix bugs, and open draft pull requests (PRs). A separate automation evaluates PRs by risk. Some low-risk changes may be merged autonomously, while others require his approval.

Agents also inspect monitoring tools such as Sentry and Better Stack, using detected issues to propose fixes. He reviews past agent conversations for recurring obstacles—such as confusing documentation, environment setup problems, or test-suite failures—and turns those lessons into updated documentation or skills.

Safety, testing, and review

Because his trading system uses real money, he emphasizes safeguards against unnoticed changes:

  • Golden-input regression tests: Fixed inputs, such as historical prices, should produce fixed outputs. He says he has thousands of tests covering trading-system configurations and critical execution paths.
  • Human-controlled production approval: Changes to protected trading paths cannot be deployed without approval from a separate GitHub administrator account, accessible only on his phone.
  • Behavioral validation contracts: YAML or plain-text descriptions specify what a workflow should do, how to check it, and how often to verify it. Monitoring agents check whether those behaviors continue to hold—for example, whether comments are collected, stored, and passed through sentiment analysis.
  • Adversarial plan reviews: Before implementation, he asks fresh agents to challenge plans, identify omissions, and raise questions. He finds the first two or three review rounds especially valuable, though he may use up to ten iterations for critical work.
  • PR checks: He describes requiring two agents to review and comment on a PR, with one acting as a critic who actively looks for weaknesses and unintended effects.

Tools and agent management

  • He describes using an official Codex app server to expose actions such as listing, creating, renaming, or archiving chats. Other agents, including Claude or Kimi, can use it to manage Codex work and identify blocked or stale tasks.
  • He uses Hermes as a “chief of agents,” including through Telegram voice notes. This lets him assign work across machines while away from his desk.
  • He also discusses voice and computer-use features: speaking to an orchestrator, controlling a computer remotely from a phone, and launching multiple subagents. He says these are useful for planning, debugging, and hands-free task management.
  • The subtitles mention several model and tool names, including Codex, Claude, Kimi, Hermes, OpenRouter, and skills such as “Superpowers.” Some specific model names may be inaccurate because the captions are auto-generated.

Principles and practical guidance

  • Keep systems as complex as necessary, but as simple as possible. The presenter warns that agents tend to add unnecessary code and complexity, increasing maintenance work, context load, and failure points.
  • Improve plans before coding. Iterative planning and fresh, independent reviewers can expose gaps early and reduce rework.
  • Make agent output easier to act on. He is exploring concise “solution packages” that offer a few concrete options with examples rather than lengthy explanations.
  • Recognize the human as the bottleneck. In an always-on system, the value of agents depends heavily on the quality and clarity of the decisions they receive. He characterizes his own role as a product owner who directs work and reviews PRs rather than writing code manually.

Other ideas discussed

A swarm of agents can review old chats to extract recurring preferences and architectural decisions, then compile them into shared documentation.

He describes a possible social-media monitoring service that could alert users when specified keywords or brands appear in online discussions. His own monitoring spans many sources and supports investment research and sentiment tracking.

He reports a substantial productivity increase from these workflows, while acknowledging that some research and creative-investment tasks remain difficult to solve fully.

Main speaker and source

Nath Aston, the video’s presenter, describes his agent workflows, tools, and hedge-fund research and development setup.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video