From Rule-Based Bots to AI Agents: A Short History
Four decades of trading automation in one read: program trading, HFT, retail crypto bots, and today's AI agents. What each era made possible, what blew up, and which lessons survived.
In October 1987, automated selling programs helped erase 22.6% of the Dow in a single session. Today, you can describe a trading rule in one sentence and have software execute it while you sleep. The distance between those two moments is the evolution of trading bots: four decades of markets handing more work to machines, usually gaining efficiency, occasionally failing in public and at scale. This is that history, era by era — what each wave made possible, what it broke, and the lesson every generation seems to relearn the hard way.
Program trading and the 1987 wake-up call
The prehistory is mechanical. In 1976 the New York Stock Exchange switched on DOT, the Designated Order Turnaround system, letting brokers route orders electronically instead of walking them across the floor. By the mid-1980s, institutions were using its successors to execute "program trades": entire baskets of stocks bought or sold as one instruction.
The signature strategy of the decade was portfolio insurance. The idea was reasonable in isolation: as the market falls, mechanically sell index futures so a portfolio's downside stays capped. Then Monday, October 19, 1987 arrived. Falling prices triggered selling, which pushed prices lower, which triggered more selling from every fund running near-identical logic. The Dow closed down 22.6%, still the largest one-day percentage loss in its history.
The lesson was not "automation is dangerous." It was subtler: an automated rule that thousands of others also run stops being a hedge and becomes a crowd. Black Monday taught markets that correlated automation is itself a risk factor, and that any rule needs a cap on how much it is allowed to do. The circuit breakers introduced in its aftermath are exactly that idea, applied market-wide.
Electronic markets and the speed race (1990s–2000s)
The next two decades moved trading from floors to servers. Electronic communication networks matched buyers and sellers directly, online brokers put order entry in living rooms, and decimalization in 2001 shrank the minimum US price increment from a sixteenth of a dollar to a cent, compressing spreads. By the mid-2000s, high-frequency firms were paying to place servers meters from exchange matching engines, competing in microseconds.
Retail traders were bystanders in that race but beneficiaries of its side effects: tighter spreads, near-instant fills, commissions trending toward zero. The era's failures, though, previewed everything that matters about automation at speed.
On May 6, 2010, the flash crash took the Dow down roughly 1,000 points in minutes before prices mostly recovered within half an hour. Automated market makers withdrew simultaneously as volatility spiked; with resting bids gone, some large-cap stocks briefly printed at a penny. Two years later, on August 1, 2012, Knight Capital deployed new software to seven of its eight trading servers. The eighth ran stale code that an old flag reactivated, spraying errant orders across roughly 150 stocks. The firm lost about $440 million in under an hour and did not survive as an independent company.
Two lessons carried forward. Speed multiplies error faster than it multiplies profit, so kill switches and pre-trade risk checks stopped being optional. And the latency game closed permanently to individuals — the useful inheritance for retail was never speed, but the discipline invented to contain it.
The 2010s: the evolution of trading bots reaches retail
Automation first reached individuals through FX, where MetaTrader's "expert advisors" let anyone run scripted strategies on a retail broker account. The real explosion came with crypto. Exchanges there launched API-first, never close, and settle in minutes: a native habitat for software. The 2017 bull run minted a cottage industry of bot platforms offering grid bots (a ladder of buy and sell orders across a range), scheduled DCA bots, and signal-following automations, no code required.
For the first time, an individual could run genuine 24/7 automation on a live account. Many still do; the retail-investor segment of algorithmic trading was estimated at roughly $3.95 billion in 2025, growing around 13.4% annually (Grand View Research, 2025). The mechanics of automated trading had become a consumer product.
The blow-ups changed shape accordingly. Expert advisors were sold on curve-fit backtests that collapsed live. Martingale bots, which double position size after each loss, printed steady small wins until one strong trend erased entire accounts. And a generation of users learned API key hygiene the hard way, granting withdrawal permissions to services that later vanished. The era's lesson: access was solved, judgment wasn't. A tool being real does not make its marketing honest.
From ML signals to LLM agents (2018–today)
Machine learning entered quietly. Institutions used it for signal research and execution optimization through the late 2010s, while much retail "AI" was a label stuck onto ordinary grid logic. The substantive shift came in 2023, when large language models made a new interface possible: software that reads a sentence, or a news wire, and converts it into structured, monitored conditions.
That is the real difference between this era and the last. An AI agent differs from a classic bot less in execution than in input: fixed if-then rules versus the ability to interpret intent and unstructured data, then act under human-set limits. It is the perceive-reason-act pattern that defines agentic trading.
The research community has already delivered this era's cautionary findings, for once before the blow-up rather than after it. Backtests of LLM strategies are often inflated by memorization: a model whose training data overlaps the test window can "remember" what happened rather than predict it (arXiv:2505.07078, 2025). And a 2026 audit of 77 studies on LLM trading agents found little support for lasting outperformance, placing the demonstrated value in execution discipline, monitoring, and research support instead (arXiv:2605.19337, 2026). Serious platforms in this wave, Obside among them, respond with an architectural split: the language model interprets what you mean, while deterministic engines handle prices, math, and order routing.
Four decades, one recurring lesson
| Era | What became possible | What blew up | Lesson that stuck |
|---|---|---|---|
| 1980s program trading | Basket execution at machine speed | Black Monday's feedback loop (1987) | Correlated automation is crowd risk; cap what any rule can do |
| 1990s–2000s electronic markets | Cheap, instant fills for everyone | 2010 flash crash; Knight Capital (2012) | Speed multiplies error; kill switches and pre-trade checks |
| 2010s retail bots | 24/7 automation without code | Curve-fit EAs, martingale ruin, stolen API keys | Validate before trusting; never grant withdrawal rights |
| 2020s LLM agents | Strategy expressed in plain language | Inflated AI backtests, autonomy hype | Demand out-of-sample proof; keep guardrails human-set |
Read down the third column and the pattern is hard to miss: every wave rewarded the disciplined and punished the credulous, and the punishment always arrived through the era's newest capability. Portfolio insurers trusted a formula. Knight trusted a deployment. Bot buyers trusted a screenshot of a backtest. The temptation being sold changes; the failure to verify does not.
The fourth column tells a more hopeful story: the survivors' toolkit accumulated. Position caps, circuit breakers, tested rules, controlled rollouts, scoped permissions. Each era added a layer, and for most of this history the full stack existed only inside institutions. Where the current wave goes next is its own question (see agentic trading in 2026), but the toolkit's direction of travel is clear: downward, toward individuals.
What this history hands you
The honest summary of four decades is that automation kept its promises about capability and broke them about ease. Machines did take over the watching, the arithmetic, and the execution. What they never supplied is the judgment layer: knowing whether a rule deserves to run at all.
What is genuinely new right now is not intelligence. It is that the institutional discipline stack finally became accessible in plain language. On Obside, the path Black Monday's survivors would recognize is the default one: you type "buy $150 of a global equity ETF on the first trading day of each month, skip the buy if the index fell more than 8% over the prior month," the copilot restates it as exact monitored conditions, and the agent runs against years of history, then on paper with live data, before a single real order routes — with position caps and drawdown limits enforced at execution time.
History's closing argument is simple: the next blow-up will find the people who skipped the checklist. Be the other kind. Describe your rule to Obside's copilot and let it earn your trust in backtest and paper before it ever touches capital.
Educational content only. This is not investment advice. Trading involves risk, including possible loss of capital.
FAQ
The earliest ancestors were exchange order-routing systems like NYSE's DOT, launched in 1976, and the institutional program-trading systems built on top of it in the 1980s. They executed pre-defined baskets and hedging rules at machine speed. Retail-accessible bots arrived decades later, first as MetaTrader expert advisors in FX, then as crypto grid and DCA bots during the 2010s.