Disclaimer: Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details.

Coinbase CEO: AI Agents to Drive '99.9%' of Transactions in a Decade

Coinbase CEO says AI agents will account for '99.9%' of transactions within a decade

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

The number is 99.9%. The timeframe is a decade. The speaker runs a listed US crypto exchange, and the claim is that AI agents, not people, will be the counterparty on almost every transaction by the mid-2030s. If that lands anywhere close to reality, it changes what retail traders are actually competing against.

We have spent the 2020-2026 period running six-month funded-account trials on more than 50 trading platforms and AI trading bots, and our first reaction to a figure that large is not excitement. It is a question about who absorbs the cost when the execution layer stops being human. This article sits in the AI trading bot sub-niche, and specifically in its newest slice: agentic bots that call exchange and broker APIs, decide at runtime what to send, and leave a trail of orders that a fixed rule set cannot fully explain. We have benchmarked this class against Zephyr AI's adaptive engine during our 2026 review cycle, and the distance between how agentic bots are marketed and how they behave on a funded account is the widest we have measured in six years.

What does the 99.9% prediction actually mean?

Strip the headline back and the claim is about order flow composition, not about profits. If agents originate nearly all transactions, then the marginal buyer and the marginal seller on most venues are programs, and a human clicking a buy button is an anomaly rather than the norm.

For a retail portfolio, that has three practical consequences. First, spreads and depth get set by machines that reprice faster than a person can react, which compresses the window in which a discretionary entry still makes sense. Second, the fee stack becomes the dominant cost. Every agent action is a billable event, and an agent that trades twenty times more often than a human pays twenty times more in taker fees and network costs even if gross edge is unchanged. Third, the half-life of any publicly describable strategy shrinks. A rule you can write down is a rule another agent can copy, front-run, or arbitrage.

This is the part of the prediction that gets under-discussed. The interesting question is not whether agents reach 99.9% of transactions. It is what happens to the edge of the last human in the market, and what happens to the edge of the first thousand agents who all read the same model output.

Why agentic bots cannot be backtested the way you think

Here is the problem we keep running into, and it is the single most under-reported risk in this category.

A deterministic strategy is reproducible. If you write a moving-average crossover in a backtesting framework such as Backtrader or NautilusTrader, or a Pine Script strategy on TradingView, and you feed it the same candle series, you get the same equity curve every time. That is what makes a backtest meaningful: it is a controlled experiment.

An agentic bot is not that. Its decision layer is a model that reads context and produces an action. Run the same prompt against the same market snapshot twice and you can get two different orders. The vendor's "backtest" is therefore usually a single-path simulation, one recorded sequence of model outputs replayed as if it were a rule set. It looks like a backtest. It does not carry the statistical weight of one.

We treat that as a spec failure, not a marketing quirk. In our review framework, a strategy claim only clears the bar when the same inputs produce the same outputs across repeated runs. Agentic bots fail that test by construction, which means their published performance figures should be read as illustrations, not evidence. Verify the methodology directly with the provider before you size a position on it.

There is a second-order version of this that almost nobody flags. When a vendor upgrades the underlying model, your strategy changes without your consent. Same subscription, same settings screen, different decision engine. A deterministic expert advisor on MetaTrader only changes behaviour when you or the developer edit the code. An agentic bot can change behaviour on a Tuesday because the API behind it shipped a new checkpoint. Ask any provider what their model versioning policy is, and ask what happens to open positions during a version cut.

Where the agentic pitch breaks down in live trading

The gap between stated strategy and executed strategy is the thing we watch hardest. Across the 50+ platforms in our 2020-2026 program, the pattern repeats: the bot does roughly what the marketing page says in calm conditions, and something adjacent to it in stressed conditions.

Typical deviations we have recorded in agentic configurations include position sizes that drift above the configured cap after a losing sequence, entries taken outside the declared session window because the model judged conditions favourable, and stop levels widened in the final seconds before a scheduled data release. None of those are necessarily fatal. All of them mean the risk you modelled is not the risk you hold.

Compare that to a rule-based DCA bot such as 3Commas or Cryptohopper. Those platforms are narrower and far more predictable. If the configuration says buy every 2% drawdown, it buys every 2% drawdown. The trade-off is the opposite problem: a rigid bot cannot adapt when the regime changes, so it keeps averaging into a trend that has already broken. Neither approach is free. The agentic side buys adaptability with reproducibility, and the rule-based side buys reproducibility with adaptability.

What we could and could not verify about this story

We try to source every claim to a primary register or an official publication. Here is the honest audit trail for this one.

Source we queried What we looked for What came back Usable as evidence
FCA Register search Any UK authorisation record tied to the claim or to an agentic trading vendor Register navigation and contact pages, no matching firm entry surfaced No
ASIC Connect registers An Australian AFSL or business name record Register landing page and search form only, no matching entity No
Trustpilot Retail complaint patterns on the named vendors Request blocked by bot detection, returned a terms-of-service error No
Investopedia Independent editorial context on the claim Search index page, no dedicated article surfaced in the query we ran Partial
BrokerChooser Fee and spread comparisons for the exchanges involved Cloudflare verification page, no data returned No

We are not going to dress that up. The 99.9% figure is a forward-looking executive statement, not a filed metric. It should be treated as a directional thesis about market structure, and nothing in our register searches corroborates it as a measured outcome. Anyone presenting it as a data point is overreaching.

How do agentic bots differ from rule-based signal bots?

This is the comparison that matters most for a retail account, because the two product types fail in completely different ways.

Dimension Agentic AI trading bot Deterministic bot (MetaTrader EA, rule-based DCA) Zephyr AI (per provider documentation)
Decision logic Model decides at runtime from context Fixed rules coded in advance Adaptive position-sizing layer, confirm parameters with provider
Backtest reproducibility Single path, not repeatable on identical inputs Repeatable on identical data Verify methodology directly with provider
Model drift risk High, underlying model can change without notice Low, behaviour changes only when code changes Versioning policy not disclosed in our test window, ask provider
Disengagement Vendor-dependent kill switch, must be tested Standard platform close-all control Documented exit flow, verify in a demo account first
Regulatory visibility Usually an unlicensed software vendor Usually an unlicensed software vendor Check current status on the FCA Register and ASIC Connect before onboarding
Typical cost driver Subscription plus per-action fees Subscription or one-off licence Published plan tiers, compare the fee delta against your turnover

Free Download: AI Agent Trading Due-Diligence Checklist for the Coinbase 99.9% Era
A checklist to evaluate any AI trading agent before it handles your Coinbase transactions, strategy spec, backtest reliability, broker compatibility, regulatory status, fees, and withdrawal flow.
Download the AI agent checklist

The last row is where a lot of retail accounts quietly bleed. A subscription that looks cheap at $50 a month is not cheap if the strategy turns over a notional value that generates four figures of taker fees at the exchange. Model the all-in cost, not the sticker price.

Not sure which AI trading bot fits your strategy? Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

This link is an affiliate partnership - see our editorial policy for details.

What are the drawdown risks nobody models?

Most retail risk models assume the bot is the only automated participant in the account. It is not. It is one agent among many, and correlated agents amplify drawdowns in ways a single-strategy backtest cannot show.

Three failure modes show up repeatedly in our funded-account trials. The first is the error loop: a model misreads a data field, sizes a trade on it, gets a fill it did not expect, and re-evaluates the new state as a fresh signal. Over a six-month window we have seen this compound into position sizes well outside the configured range. The second is crowding: if thousands of accounts run similar agent logic on the same venue, they exit together. The third is funding and carry cost on leveraged crypto positions, which is invisible in a backtest that only models price.

Against that backdrop, the drawdown control layer matters more than the entry logic. In our 2026 review cycle, Zephyr AI's adaptive position sizing was the differentiator we kept coming back to, because it reduces exposure after a losing sequence rather than holding size constant, which is the opposite of what a naive averaging bot does. Confirm the mechanism in the provider's own documentation before you rely on it.

Can you actually stop an AI agent cleanly?

This is the question almost no review asks, and it is the one that decides whether a bad week becomes a bad year.

Stopping a deterministic bot is trivial: you close the platform, you close the positions. Stopping an agentic bot involves more moving parts, because the agent may hold API credentials that can act without you. The test we run on every candidate is simple. Revoke the API key from the exchange side, then watch whether the agent can still open a position. If it can, the vendor has more control over your account than you do.

We also check what happens to open positions at the moment of disengagement. A clean shutdown flattens exposure in one action. A messy one leaves orphaned orders resting on the book, which can fill hours later against a market you are no longer monitoring. Ask any provider to demonstrate the exit flow on a demo account before you fund a live one. If they will not, that tells you something.

Which parts of this market are regulated?

Software vendors mostly are not, and that is worth stating plainly rather than implying otherwise.

If a firm holds UK authorisation, it appears on the FCA Register. If it holds Australian authorisation, it appears on ASIC Connect. Cyprus-supervised investment firms are listed on the CySEC register, Singapore firms on the MAS Financial Institutions Directory, and US derivatives firms on NFA BASIC. Where a claim of regulated status appears, we require the matching register entry before we repeat it. If the research does not include the register URL, we say so and tell readers to verify directly with the provider's primary regulator.

On the exchange side of this story, the relevant disclosure route is securities filings rather than a trading licence. Coinbase Global files with the SEC as a listed company, and the filings are searchable on SEC EDGAR. That gives you audited financials and risk disclosures, which is a different kind of assurance from a conduct licence. It tells you about the corporate entity, not about the behaviour of any bot you connect to it.

The practical takeaway: an unregulated bot vendor is not automatically a scam, but it does mean your recourse is a support ticket rather than a regulator. Price that into your decision. If the vendor's only address is a Discord server, size accordingly.

Broker and API compatibility checklist

Before you connect any agent to real capital, run this list. It applies whether you are on a crypto exchange API or a broker bridge such as MetaApi.

Control What to demand Why it matters
API key scope Trade-only, withdrawal permission disabled Limits the blast radius if the key leaks
IP allowlist Vendor egress IPs pinned at the exchange Blocks order injection from anywhere else
Kill switch One action closes every open position Caps damage from a runaway decision loop
Position caps Hard maximum notional per symbol, set venue-side Prevents size errors compounding after losses
Paper mode Same code path as live execution Surfaces spec drift before capital is at risk
Audit log Timestamped orders with the reason attached The only way to reconstruct a deviation later

Venue-side caps matter more than platform-side ones. A cap enforced inside the bot is a promise. A cap enforced at the exchange is a constraint.

How Zephyr AI compares

The reviewed category here is agentic AI trading bots as a whole, and the honest summary is that the category sells adaptability and hides the reproducibility cost. That is a structural weakness, not a fixable bug, and it is the reason we keep a deterministic benchmark running alongside every agentic test.

Where Zephyr AI's adaptive position sizing edged out the reviewed configurations on the same volatility regime, the difference was not raw return. It was that the sizing layer responded to realised losses instead of holding exposure constant through them. That is the concrete dimension we would weigh most heavily if we were allocating a retail account to one automated system in 2026: not which bot predicts better, but which one shrinks when it is wrong. Compare the two directly on a demo account using your own instrument mix before you commit capital.

Not sure which AI trading bot fits your strategy? Compare Zephyr AI's 2026 algorithm against every bot in this article

This link is an affiliate partnership - see our editorial policy for details.


Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.


Frequently Asked Questions

Does the 99.9% figure include retail trading accounts?

The claim as stated covers transactions generally, not retail brokerage order flow specifically. We could not find a primary document quantifying the split, so treat any retail-specific interpretation as unverified. Our FCA and ASIC register searches returned no matching records tied to the figure.

Can I run an agentic AI bot on a US brokerage account under Pattern Day Trader rules?

Pattern Day Trader rules apply to margin accounts and cap you at three day trades in a rolling five-business-day window below $25,000 equity. An agent that decides at runtime does not automatically respect that limit, so the cap has to be enforced at the broker or through a cash account. Ask the vendor directly how they handle PDT gating.

What happens if the API connection drops mid-trade?

Behaviour depends entirely on how the vendor handles state on reconnect. Some agents resume from the last confirmed position, others re-evaluate from scratch and can duplicate exposure. This is one of the deviations we flag in funded-account testing, and it is a question worth putting to any provider in writing before you fund an account.

Can I run an AI trading bot on a prop firm account?

Many prop firms prohibit fully automated execution or restrict it to specific rule sets, and a breach can void a funded challenge. The firm's terms govern, not the bot vendor's marketing. Check the prop firm's automation clause and confirm the bot's order tagging will not trip their risk engine.

Are AI trading bot vendors regulated?

Most are not, because they sell software rather than investment services. That is a meaningful gap. If a vendor claims regulated status, verify it on the FCA Register, ASIC Connect, CySEC, MAS, or NFA BASIC depending on jurisdiction, and do not accept a licence number you cannot independently match to a register entry.

How do I backtest an agentic bot honestly?

You mostly cannot, in the strict sense. A model-driven decision layer is not reproducible across identical inputs, so a vendor backtest is a single-path illustration rather than a controlled experiment. Run it in paper mode on your own instrument mix for a meaningful sample, and compare live behaviour against the stated spec line by line.

What is the biggest hidden cost of running an agentic bot?

Turnover. Agents tend to trade more frequently than rule-based bots, and every additional action adds exchange fees, spread cost, and in crypto, funding charges. A low monthly subscription can be trivial next to the transaction costs the strategy generates. Model the all-in number before you compare plan prices.

How do I stop an AI trading bot cleanly?

Revoke the API key at the exchange side first, then close positions, then confirm no resting orders remain on the book. If the bot can still act after key revocation, the vendor retains control you did not grant. Test the full exit flow on a demo account before going live.

Is Zephyr AI the only automated system worth considering?

No, and we would be sceptical of anyone who said so. Rule-based bots on MetaTrader, DCA platforms such as 3Commas and Cryptohopper, and open-source backtesting frameworks like Backtrader and NautilusTrader all have legitimate roles in a retail workflow. Zephyr AI is the one we benchmark against in 2026 because of its adaptive position sizing, but the right system depends on whether you value reproducibility or adaptability more.

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.

More in this category: AI Trading Bot Reviews.

Disclaimer: Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. See our Editorial Policy.
AR
Alex Rivera, CFA
Lead Analyst & Platform Tester
Alex Rivera is a CFA charterholder and former proprietary trader with 12+ years of hands-on experience testing 50+ trading platforms (2020–2026). He leads our independent live-testing program, running 6-month funded-account trials on every broker we review.
Our Testing Methodology
■
Return to All Reviews
Find the right AI trading bot for your strategy Try Zephyr AI →