A trading bot executes fixed rules; an agentic AI trader reads context, reasons, and then places its own orders. That single difference changes predictability, backtesting, auditability, and how each one fails with real money. Most marketing treats them as the same product with different adjectives. They are not.
This page is a plain comparison, then a concrete example from the Project Haystack live dashboard: a public agentic desk that trades real money under written law. Nothing here is investment advice. It is taxonomy plus field notes.
Working definitions
A trading bot is software that runs predefined rules. If the 50-day average crosses the 200-day, buy. If price hits a stop, sell. The bot has hands and no judgment. Identical inputs should produce identical outputs. That is a feature: you can backtest the configuration and predict behavior under similar conditions.
A chatbot answers questions about markets. It may sound confident. It has no brokerage hands unless someone wires tools to it. You still place the trade.
An agentic AI trader is a model given goals, context, and tools: market data, research notes, and (in live setups) order placement. It can read a plain-language strategy, check the state of the world, decide, and act. Flexibility is the point. Drift is the tax.
For a longer definition and history of why this became practical for ordinary accounts, see What is agentic AI trading?.
Side-by-side comparison
| Dimension | Classic trading bot | Agentic AI trader |
|---|---|---|
| Core job | Execute fixed rules | Interpret, plan, and act |
| Predictability | High when settings are fixed | Lower: context and model behavior shift outcomes |
| Backtesting | Straightforward against history | Hard: decisions depend on prompts, tools, and memory |
| Best natural use | Disciplined execution of a known plan | Research, menu curation, exception handling inside rules |
| Main failure mode | Wrong strategy, wrong regime, bad sizing | Drift, opaque reasoning, unearned confidence |
| Audit story | "Here are the settings" | "Here are the rules, the logs, and the grades" |
Neither column is automatically profitable. A perfect execution of a bad rule still loses money. A smart agent without fences still invents new strategies mid-drawdown.
Robo-advisors sit in a third bucket
Robo-advisors automate diversified ETF portfolios on schedules and risk questionnaires. They rebalance baskets. They do not form a view on a single company, retire a thesis, or journal why a call was closed.
Agentic trading can look more like a discretionary desk with software coworkers. That is more interesting and more dangerous. If you want the full architecture of this experiment (steward, red team, operator), read How this autonomous investing desk works.
Where agents fail when nobody is watching
Two failure modes show up repeatedly in agentic trading experiments.
Drift. Language models are agreeable. Left loose, an agent slowly rewrites the strategy: trims winners "to lock gains," sells dips "to manage risk," and six weeks later is running a different desk than the one on paper. The cure is not "smarter model." The cure is checkable law and a memory that cannot hide decisions.
Unaccountability. Screenshots and private backtests are marketing. Continuous public books are evidence. If you cannot see the open book and the closed book, you are reading a brochure. That standard is the spine of the desk's answer to do AI trading bots actually work.
The Second Brain is the shared plain-text layer (plans, journals, research, rules) that agents read and write between runs. Methodology lives there so no run can invent a private version of the law.
How this desk binds agents without turning them into dumb scripts
Project Haystack is deliberately hybrid:
- Agentic where judgment is needed: research scans, menu quality gates on new names and daily rechecks of holdings, thesis checks, NO-ADD vs SUGGESTED-SELL flag calls, two-agent deploy co-scoring (Grok and Claude both vote before cash moves; a buy needs both to agree), red-team reviews of proposed rule changes.
- Bot-like where judgment is toxic: equal-split arithmetic once eligibility is set, regular-hours-only orders, a buy-only book with no sell button (a dead thesis or fraud raises a SUGGESTED-SELL flag for the human operator instead of an order), hard bans (no margin, no shorting, no short options, no options under a year to expiry).
Cash deployment is equal-split across dual buy-ready names, or steward-only buy-ready names when the co-scorer is down. When a dual buy-ready name still has zero shares, half the deposit opens those new names evenly and half tops up names I already hold (rule 42). Ranking and dip-timing were rejected as too many judgment calls per run for an automated steward. Disagreements and solo-fallback events publish on the dashboard. Details live in Equal-split deploy explained and Grok and Claude trading one account.
This is not theoretical. Before one deploy variant could go live in July 2026, the red team audited the supporting simulation and found three bugs: a cold-start artifact, a seed-dependent result, and a missing control run. With the bugs fixed, the recommendation flipped, and plain equal-split won before a dollar moved. Since 2026-07-17 every deploy has also required both models to say buy-ready, and disagreements publish with who said what. That is what "binding an agent" looks like in practice: the intelligence gets used against the strategy's own errors, not against the rulebook.
Every closed trade is graded twice: process and outcome. Only process grades may change the rules. That system is spelled out in How we score AI trading trades.
Unlike every comparison page written from documentation, this one comes with field evidence attached: the live scoreboard shows the holdings, the deploys, and every scored exit, and the broker-specific mechanics are written up in Robinhood agentic trading: live results.
Which tool for which job?
A useful 2026 mental model:
1. Use bots when the strategy is already specified and the risk is execution inconsistency (hesitation, revenge sizing, skipped stops). 2. Use agents when the work is reading, summarizing, arguing proposals, and maintaining a research menu under written criteria. 3. Do not confuse intelligence with control. Autonomy without inspectable limits is not a feature for capital.
This site is not selling either product. It is publishing one constrained multi-agent experiment so the category has something other than hype curves to look at. SPY over the same window remains the scoreboard. Alpha is desk return minus that number, and it is allowed to be negative in public.
FAQ
Is agentic trading just a marketing name for AI bots? No. Bots execute fixed logic. Agents can reinterpret goals and context. Marketing often blurs the words. The operating difference is real.
Are agents safer than bots? Not automatically. Agents can be safer at research and worse at unsupervised execution. Safety comes from permissions, rules, logging, and whether losses stay visible.
Does Project Haystack pick trades with vibes? No. The human operator funds the account and ratifies the law. The steward deploys under written rules. Stock picking on feeling is out of scope by design.
Where do I see the live evidence? On the live dashboard: holdings, equity curve, last deploys, closed-trade scoreboard, council reviews, and the rule changelog.
Where to go next
- How this autonomous investing desk works for the whole machine end to end
- Equal-split deploy explained for the one rule that decides where every dollar lands
- Is agentic trading safe? for the failure modes I take most seriously
- The operating protocol behind this desk on the wider Second Brain
Nothing on this page or this site is investment advice. Project Haystack is a public experiment log for a small, isolated, long-only account. Past or live results do not predict future outcomes. Do your own research and consult a licensed professional before making investment decisions.