Stardust - an experiment in agentic stock trading

I saw nof1.ai before I saw a reason to build this — a leaderboard pitting different models against each other, each handed real money to trade crypto. My friend Manan sent it over, and the question stuck: can these models actually generate money, or does agentic trading fall apart the moment fees, slippage, and real order books get involved? We turned it into a friendly competition — who could build the better system for the Indian market, not crypto.

Getting there took three attempts at the framework, not the strategy. I ran versions on OpenClaw, on Pi, and on others, largely to learn agentic systems thinking by building one badly first. Pi came closest — a minimal, highly extensible core with real self-learning loops — but it's CLI-first: no clean way to run a web dashboard or a Telegram bridge, and it fumbled restarting into a fresh daily context. A nudge from another friend, Nikhil Jois, pointed me at Hermes. I scrapped the Pi build entirely and rebuilt from scratch on Hermes — better stability, better memory, better self-learning, full stop.

Stardust went live on Dhan on 18 August 2026. My existing portfolio stayed ring-fenced at first while the agent found its feet — through a static-IP outage, a timezone bug, and a security scare, in roughly that order. Once it held up, I opened the whole account to it on 1 September: ₹30,000 of real capital to start, ₹90,000 under management as of this writing. None of it is fake — including the roughly $0.70–$1.00 a day in model tokens and the ₹500/month Dhan API subscription it costs just to keep the lights on.

I'm publishing the design, not selling a system, because I'd rather have people find the holes in it than find them myself, eventually, the expensive way. If this pushes even a few more people to experiment seriously with agentic trading — and send real scrutiny back my way, or better, real improvements — the writeup did its job. It's still a heavy lift: constant configuration, constant testing, real money at stake on both sides of the ledger. Nobody should copy this and walk away.

The LLM never talks to the broker directly, and never touches Python it wasn't handed. It reasons in language, then calls a named tool; the tool shells out to the same tested CLI a human would run by hand. That's a deliberate design choice, stated plainly in the agent's own operating manual: the agent figures out what to do and calls the tools, instead of running raw scripts. Every action — a scan, a buy, an exit, a token refresh — passes through the identical, already-hardened code path regardless of who or what initiated it.

It's also why there's no MCP server here. Dhan's API is already just Python functions the engine calls; wrapping that a second time behind a protocol server would mean two credential surfaces, two places for IPv4/whitelist quirks to diverge, two things to keep in sync. Instead the plugin registers native tools that shell into the same venv, the same scrip-guard, the same state file the CLI has always used. One environment, reused, not re-implemented.

There's a strict division of labor underneath that: a skill teaches the model how to think — the mandate, the invariants, the accumulated pitfalls — and a tool is the only thing it's allowed to act with. Every unit of real work — a quote, a scan, an order, a ledger write — exists as a named tool, never as prose telling the model to run a script from memory. The skill can reason all it wants; it still has to reach for a tool to actually do anything.

The premarket stage doesn't only look at price. It reads the day's news through the same scan pipeline and folds that straight into both the sized "setup of the day" and the standing watchlist — a name can get added or dropped purely off an overnight headline, before a single candle of the new session has printed.

The cadence controller

The interesting piece isn't any single job — it's that CadenceSync exists at all. Every five minutes during market hours, a cheap deterministic script (no LLM call) looks at recent order activity, not position count, and decides how often the intraday agent should even bother waking up. A carried position sitting behind a live broker-side stop doesn't need a fast poll; a name that just got bought or sold does.

The base system prompt (SOUL.md) is deliberately generic — a short, stock identity: "you are Hermes Agent, helpful, direct." It says nothing about trading. Every actual guiding principle lives one level up, in the skill that gets loaded fresh per cron job — which is also why those principles can evolve daily without ever touching the agent's core identity.

Structurally, SKILL.md opens with a mandate ("maximize long-term profit, compound — capital is always pulled live from the broker, never hardcoded"), a tool reference table so the model knows what its hands are before it reasons about what to do with them, and a set of hard invariants that hold regardless of what the LLM concludes in any given session.

The most transferable part, though, is the long pitfalls section — real production mistakes, each rewritten generally enough to prevent the whole class of bug, not just the one instance that happened. A representative excerpt of the kind of thing that lives there:

On scope, deliberately: Stardust currently trades cash equity only — no F&O, no global markets. That's not a technical ceiling, it's a sequencing choice: prove the tools and the strategy hold up in the simplest, most reversible instrument before handing the agent anything with leverage or a different risk shape. Access widens in stages, and only after the stage before it has actually been trusted with real capital for a while.

Worth saying plainly, too: this is not a high-frequency system. The fastest wake interval anywhere in the schedule is 10 minutes, gated behind a cadence controller that actively tries to slow things down, not speed them up. Nothing here is competing on latency — it's competing on not doing anything stupid between wake-ups.

Reconciliation isn't a nightly cleanup job, it's a rule enforced at every stage: reconcile compares state against the broker's positions, holdings, funds, and ledger, and never trades — it only flags drift. A recurring class of early bugs came from treating /positions (today's fills only) and /holdings (delivery carried from prior days) as interchangeable; they aren't, and the engine now asks each endpoint only what it actually knows.

Cost accounting is kept separate from trading performance on purpose — LLM inference spend and broker fees are tracked against a GROSS/NET line, not folded into strategy P&L, so a good trading day and a good token-cost day are never confused for each other. Market data for the scanner leans on free sources (NSE bhavcopy, a news scraper) rather than the paid Dhan data feed, keeping the technical-analysis pipeline independent of the flakier paid endpoint.

Three of the agent-driven cron jobs started failing intermittently with Blocked: prompt matches threat pattern 'deception_hide' — Hermes's own prompt-injection scanner refusing to run the assembled prompt before the model ever saw it.

The honest read: Stardust beat both benchmarks in every clean week, and did it by going up while the market went down — not by losing less. Over the month that compounds to roughly fifteen points of outperformance.

But the trend inside that number matters more than the number. The weekly edge has gone +3.1pp → +2.9pp → +0.8pp. In the most recent week the market was essentially flat at −0.22% and Stardust's +0.62% barely cleared it. Whatever the agent was exploiting in late August, it is exploiting less of it now. One month is far too short a window to call that either decay or noise — which is exactly why it's published rather than waited out.