The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Victor Ike’s account of building Atlas shows why getting AI-generated code to run was only one part of the job. Across three rebuilds, he encountered a poor-fit project structure, broker orders that did not match the platform’s records, hard-coded assumptions, and a planning process that grew unwieldy. His account’s central lesson is that the agent helped implement a system, but Ike had to supply the trading decisions and define how it should behave when facts were uncertain.
Why Atlas had to be rebuilt three times
Ike began Atlas on July 31, 2026. In his September 28 account, he says the first version lasted eleven days and about 117 commits before he stopped. The agent had selected a folder structure he found difficult to follow, so he restarted using an organization he had chosen himself. That was not simply a code-generation failure: the project’s initial structure and the way he had set it up did not fit how he wanted to work.
The second version ran for about a month and reached paper trading. That made it possible to test the connection to a broker, but it also exposed a more consequential problem than untidy code: the software’s understanding of an order did not reliably match what happened at the broker.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Around September, Ike started a third version, atlas-next. At first he carried assumptions from the previous design into the new one. The work expanded to 27 workstreams; at one point, he says, the repository contained about 41,000 lines of documentation alongside 18,000 lines of code. On September 23, he reset the process and switched tools. He reports that planning improved, rewrites became less frequent, and the agent asked questions or checked in rather than guessing when a decision was needed. Ike also disclosed that Claude helped write his September 28 post, so his assessment of that workflow should be read with that context.
#1 Best Overall
What went wrong when Atlas reached paper trading?
An order missed its take-profit
During a practice trade sent to OANDA, the order arrived without its take-profit instruction attached. Ike says the strategy represented in the order also differed from the one he had described. When the broker’s order could not be matched to Atlas’s own records, the platform could not account for the position; after Ike closed it manually, the software could not account for that close either.
The trade lost money, but the more important failure was operational: the expected risk instruction was missing, and the platform’s internal state did not reflect the broker’s. Ike reports that a later paper order included its take-profit and reconciled correctly, although that trade also lost. One successful reconciliation does not establish that the system was reliable or that its strategy had an edge.
Assumptions and process also accumulated
Ike says EUR/USD and OANDA were hard-coded throughout the second version, making the design less portable. He also describes versioning and documentation practices that assumed a maturity the project had not reached, documentation that went stale, the agent no longer following earlier directions, and weak historical backtesting. These issues reflect a combination of system design, project setup, and tool behavior; his account does not establish that the model alone caused them.
What changed in the third version’s design?
In a September 25 description of Atlas, Ike outlined a more explicit approach to evidence and failure handling. He said stored price data served as fixed input for deterministic backtests, and strategy settings were versioned so a result could be traced to the settings that produced it. He also described treating an order request as distinct from an open position: the platform would count a position only after the broker confirmed a fill.
Rank #3
In that design, unclear order outcomes blocked new entries. Stale prices or disagreement between Atlas and the broker halted a run, and the live-server connection was disabled in that version. These are safeguards Ike described, not independently audited behavior or proof that every failure case was covered. Their significance is the principle behind them: when the platform cannot establish what the broker did, it should not confidently act as though its own records are correct.
What the agent could not decide for the trader
Ike’s experience supplied requirements that were specific to trading rather than ordinary application development. He had to decide what to do with a setup when the market closed for the weekend, how long to wait after reopening before trading resumed, and what qualified as a valid trade setup. He wrote that “Someone building a trading platform without that experience wouldn’t know to ask, and an agent won’t ask on its own.” That is his judgment based on this project, not an independently tested rule about every AI agent.
His division of labor was to make those domain choices and use the agent to implement them. The work therefore depended on more than translating a request into code: Ike needed to specify the strategy, the market’s operating conditions, and the response to uncertain broker state. The account also notes a 30-hour Udemy course on writing forex trading algorithms in Python, underscoring that learning the subject was part of his own effort rather than a capability the agent supplied.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWhat “working” meant at the September 28 update
By the September 28, 2026 publication point, Ike said version three ran deterministic backtests on stored data and that he had begun a paper run on his OANDA practice account on September 27. He explicitly said Atlas was not ready for real money. Those dated status claims do not establish later performance, live deployment, profitability, or whether paper results would carry over to a live account. A backtest or a paper order can show that software executed a particular process; neither, by itself, proves a strategy is profitable or safe to automate.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What this account says about building with AI agents
The three iterations are a case study in where implementation assistance ends and system responsibility begins. An agent can help produce code, but the builder still has to make the architecture comprehensible, define trading behavior, decide what evidence counts as a fill, and specify what the software must do when its records and the broker disagree.
Ike’s experience does not establish a universal ranking of AI coding tools or a general success rate for AI-built trading systems. It does show why “the code runs” is an insufficient test for a broker-connected trading platform: the intended strategy and risk instructions must survive implementation, and the platform must reconcile its state against the broker’s before it treats an order as a position.
Sources: Victor Ike, “Three rebuilds: what building a trading platform with AI agents actually took” (September 28, 2026); Victor Ike, “Building my own trading platform, three rebuilds in” (September 25, 2026).
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
