Agentic AI Trading For Beginners: A New Money-Making Era

By Kristian Fagerlie · 2026-06-04 · 7 min read

This is the question I get most: "how do I actually get started with agentic AI trading?" So I made one video that walks the whole loop, end to end, with a real $955 wallet — pick a platform, collect data, build a model with Codex, hand it to an agentic monitor, and watch the agent flip strategies mid-run when the market changes. 56 minutes, +$7 profit, full autonomous reasoning. The point isn't the $7. The point is the loop closes.

Watch the video:

Why this is the right side-hustle to learn right now

Two trends crossed in mid-2026 and they make this a uniquely good moment to start:

• Agentic AI is finally good enough. Codex 5.5 and Claude Opus 4.8 can hold a multi-step trading loop, monitor live data, and adjust position size without you babysitting. A year ago you needed a Python backtester and a lot of patience. Now you describe the goal in plain English.
• Robinhood just launched agentic stock + crypto trading. The mainstream brokerages are following the on-chain platforms into AI-driven execution. That means the API surfaces are about to get a lot more standardized, and the audience of people doing this is about to explode.

The good news for a beginner: you don't need fiat / KYC / a brokerage account to start. The two platforms I use most — Hyperliquid and Polymarket — let you connect a wallet, deposit USDC, and have an agent placing real trades within an afternoon. Detailed account setup walkthroughs already exist in the Hyperliquid post and the Polymarket post; this post focuses on the agent-side workflow, not the wallet plumbing.

The framing: hybrid model + agentic monitor

Most beginner mistakes come from picking one of these and skipping the other:

• The model: the strategy itself — entry/exit rules, position sizing, stop-loss logic. Static once defined. Best built by a strong reasoning model (Codex 5.5 high or Claude Opus 4.8) chewing through real data.
• The agentic monitor: the live loop that watches the model run, evaluates the state every minute or two, and adapts on the fly when the market shifts. This is the new part. This is what makes it agentic AI trading instead of just AI-assisted scripting.

The model alone is a Python script. The monitor alone has no strategy. Together they're a complete trading agent.

Step 1 — Account, wallet, $10 test trade

I run on Hyperliquid for this walkthrough — crypto + perp market, high volatility, easy API access. $955 in the account, MetaMask connected, API keys in a .env. Full setup is in the Hyperliquid build post.

Before any of the model work, do a sanity-check trade. I dropped a beginner.md file in the project directory describing my setup and goal, fired up Codex in YOLO mode, and said:

Read beginner.md. Create a framework based on this and our .env so we can make our first trade on cryptocurrency on Hyperliquid.

Codex built the framework, verified against the Hyperliquid SDK + docs. Then:

As a test, place a $10 Bitcoin long.

Position appeared in the Hyperliquid UI in under a second. Said "good, exit trade", and out. The API is wired up. Now we can do real work.

Step 2 — Data collection

No data, no model. Hyperliquid has a documentation page describing the historical-data endpoints (candles, order books, funding rates, recent fills, account state). I pasted that page into docs/hyperliquid.md and asked Codex:

Read this documentation file. We want to trade cryptocurrency on Hyperliquid today with the goal of making $10 today. Based on your knowledge of finance, math, and trading, collect the relevant data. We need to look for opportunity.

Codex came back with a sensible plan: skip the monthly S3 archive (too coarse for same-day trades), pull recent candles + order books + funding + market context + account risk state via the live API. It wrote a small data-collection script, ran it, dumped JSON files into data/raw/books/, data/raw/candles/1h/, data/raw/funding/.

This step matters because most beginners skip it. They jump straight to "make me a strategy" and the model invents one in a vacuum. With real recent data in scope, the model produces something it can actually justify.

Step 3 — Three hypotheses, pick one

I asked Codex for three model candidates, not one:

Based on the data we collected, create three hypotheses of a model that looks promising to meet our goal today. Be analytical, use math and data, look for that edge, be creative. Multiple trades allowed. Think hard and deep. Find the three best options. You have the full budget on the account to your disposal.

Why three? Because then you can ask the model to rank its own ideas and explain why, which forces it to articulate the edge instead of just spitting out the first plausible thing. Same pattern as in the WSB persona work — rate the candidates against the goal and pick.

Codex returned:

1. Major-coin short trend basket — short BTC/ETH/SOL when 5-min EMA-20 is dropping and 1-period return is negative. Recent data was bearish, so this was the highest-ranked option.
2. HYPE mean-reversion short.
3. Solana trend as a single aggressive bet.

Picked option 1. Codex wrote the Python implementation — capping margin usage, account-collateral aware, entry triggers per the EMA conditions, hard stop loss.

Step 4 — The agentic layer (this is the new part)

If I just ran that Python script in a loop, I'd have an algorithmic trader. Useful but not agentic. The thing that makes it agentic is wrapping it in an LLM-driven monitor that can adjust the strategy in response to market regime changes.

Codex CLI has a /goal slash command — it sets a verifiable long-running task and Codex keeps working toward it across turns. I used it like this:

/goal — Run the trading script and check on it every 2 minutes. Check the logs and state. Check the signal. Update status.md. Adjust parameters if needed. Continue until the $10 profit target is hit.

The agent now runs in a sleep loop:

• Every 2 minutes (later tightened to 60 seconds), it wakes up.
• Reads the runner's state files. Checks current signal vs entry conditions. Compares balance to start.
• Decides: hold path, adjust parameters, switch strategy, exit, or do nothing.
• Updates status.md so you can read what it did without interrupting it.
• Sleeps. Repeats.

This is the same split-agent shape I documented in detail in Building an Agentic AI Trading Heartbeat That Works, just simplified for a beginner audience. The principle is identical: lightweight observation cycle, infrequent but strong decisions.

What happened next — the market changed

I left the agent running and went to the store. When I came back, this is what I saw in status.md:

Acted on the $10-today priority since the short-bull setup was repeatedly blocked by the bullish EMA structure. Current market fits pullback mean-reversion long better than short basket. Switching strategy.

This is the part I actually care about. The market regime changed between when the model was built and when conditions for execution appeared. A static script would have sat in observation mode forever waiting for short signals that weren't coming. The agentic monitor noticed and pivoted.

Result: 3x leverage long on BTC, ETH, SOL via the new mean-reversion strategy. P&L climbed to about +$12 before the exit logic fired. By the time the close orders cleared the book, the settled profit was +$6.62 — total run time 56 minutes. We hit the $10 target intra-trade; the close-execution lag took some off, which is a known limitation I'll tighten in the next iteration.

What this proves (and what it doesn't)

What it proves: a beginner can ship a working agentic trading loop in an afternoon using Codex (or Claude Code, or Open Code), with a real wallet, on a live perp market. The agent will adapt its own strategy when the market changes, without you intervening. The mechanics work in mid-2026.

What it doesn't prove: that this is a money printer. One 56-minute run with +$6.62 profit is not signal — it's a single demo. The model-comparison work in the Codex vs Claude head-to-head showed how much one mid-run intervention or one bad ticker can swing an hour. You need many runs.

Realistic side-hustle framing: $100/month or $100/week is a sensible goal if you're learning and putting in a few hours a week. $1000+/month is harder and needs either more capital, more parallel strategies, or both. Setting a small target keeps you focused on the learning loop instead of chasing variance.

What to do next

1. Pick one platform. Hyperliquid for crypto + equity perps, Polymarket for prediction markets. Don't try both at once.
2. Set up the account. Walkthroughs: Hyperliquid or Polymarket.
3. Run the $10 test trade. Confirm the API actually works end-to-end before doing anything fancy.
4. Collect real data into data/raw/. Whatever the platform's API gives you, dump it. Don't skip this.
5. Ask for 3 model hypotheses, pick one. Force the model to defend its choice.
6. Wrap it in an agentic monitor. Use /goal on Codex or the heartbeat split-agent pattern.
7. Start small. Lose small. Learn. Run it. Read the status logs. Adjust. Don't intervene mid-run.

The complete reference for everything above lives on the Agentic AI Trading pillar — start there if you want the full picture, come back to this post when you want the beginner-walkthrough version.

Live updates and beginner questions get answered in the AI_automata Discord. I'll share the beginner.md template on GitHub for anyone who wants the starting prompt I used.

Resources

• Agentic AI Trading — Complete Guide (pillar) — the full reference.
• Building a Hyperliquid AI Agent Trader From Scratch — account + API setup.
• Building a Polymarket AI Trading Bot From Scratch — alternative platform.
• The heartbeat post — split-agent architecture for the monitor.
• Codex vs Claude on Polymarket — model comparison.
• Codex vs Claude on Hyperliquid — second comparison.
• Hyperliquid — perp DEX used in this walkthrough.
• Polymarket — alternative venue.
• AI_automata Discord — beginner questions, live updates.