Building Atlas with AI agents took three attempts, repeated architectural decisions and hands-on trading judgment—not simply a prompt and a working trading system. In his September 28, 2026 account, trader and builder Victor Ike says the first two versions ran into problems ranging from an unusable project structure to a paper order missing its take-profit and failing to reconcile with the broker. By the time of that account, Atlas was in paper trading and Ike said it was not ready for real money.
Why the first Atlas build was restarted
Ike says he began building Atlas on July 31, 2026. In his account, the agent selected a folder structure that he could not comfortably follow. After eleven days and about 117 commits, he stopped that version and restarted with a structure he had chosen. Those figures describe his project, not a typical AI development timeline.
As an Amazon Associate I earn from qualifying purchases.
The episode points to a practical early decision: the person maintaining a project needs to understand its organization. An agent can produce a plausible structure, but plausibility does not guarantee that the structure fits the builder’s working habits or that they can navigate it when something goes wrong.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11What failed when Atlas reached paper trading
The second version lasted about a month and reached paper trading. A practice order sent to OANDA did not include its take-profit instruction. Ike also says the strategy used for the order differed from the one he had described, Atlas could not match the broker’s order to its own records, and the platform could not account for the position after he closed it manually. The trade lost money, but the central failures he recounts were operational: the intended instruction was missing and the platform’s state no longer matched the broker’s.
#1 Best Overall
Ike reports that a later paper order did include its take-profit and was reconciled correctly, although that trade also lost. A correctly handled order is not evidence that a strategy works; nor does a losing practice trade, by itself, explain whether an order-handling design is sound.
Problems extended beyond one order
- Hard-coded assumptions: EUR/USD and OANDA were embedded across the project, making it less adaptable to other instruments or brokers.
- Process ahead of maturity: Versioning and documentation assumed a level of production readiness that did not fit a project Ike still considered disposable.
- Stale instructions: Documentation grew out of date, and the agent stopped following earlier directions.
- Weak historical testing: Ike says the backtesting approach did not work well.
His account describes a combination of agent choices and his own setup and assumptions. The failures should not be reduced to a claim that the model alone caused them.
How the third rebuild became overcomplicated
Around September, Ike began atlas-next, initially carrying assumptions forward from the previous design. The workflow expanded to 27 workstreams. At one point, he reports about 41,000 lines of documentation against 18,000 lines of code. These are project-specific counts from his account, not measures of how AI development generally scales.
Ike says he reset the process on September 23 and switched tools. After that, he found planning improved, rewrites declined, and the agent asked more useful questions. When a decision was needed, it checked in instead of guessing. Ike also discloses that Claude helped write his September 28 post, a relevant qualification when considering his evaluation of the workflow.
What the agent could not supply: trading decisions
Ike attributes an important part of the improvement to his own trading experience. The agent did not determine what a valid trade setup meant or, on its own, supply the operating rules he considered necessary around market closures. He had to decide questions such as: “What should happen to a setup when the market closes for the weekend?” and “How long should a system wait after the market reopens before it starts trading again?” He supplied those domain decisions and used the agent to implement them.
“Someone building a trading platform without that experience wouldn’t know to ask, and an agent won’t ask on its own.”
That is Ike’s judgment from building Atlas, not an independently established rule about every agent or trading project. It nevertheless captures the division of labor he describes: implementation help did not replace defining the strategy, market behavior and risk rules.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →What Atlas was designed to do by late September
In a September 25 description, Ike says stored price data served as fixed evidence for deterministic backtests, while versioned strategy settings made results traceable. He also describes safeguards intended to keep the platform’s state from getting ahead of the broker’s:
- An order request did not count as a position until the broker confirmed a fill.
- An unclear order outcome blocked new entries.
- Stale prices or disagreement with the broker halted a run.
- A live-server connection was disabled in that version.
These are descriptions of the design from Ike’s own posts, not the findings of an independent audit. At the September 28 publication point, he 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 also said it was not ready for real money. The account establishes no later performance, live deployment, profitability or readiness.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the three attempts show about building with an agent
Across Ike’s account, the most consequential problems were not just whether generated code could run. They concerned whether the architecture remained understandable, whether orders carried the intended instructions, whether the platform could reconcile its state with the broker, and whether the strategy’s assumptions matched real market conditions.
| Area | What Ike’s account describes | Why it mattered in Atlas |
|---|---|---|
| Architecture and maintainability | The first structure did not work for Ike; he restarted with a structure he chose. | The person responsible for the project needed to navigate and maintain it. |
| Order and broker state | A paper order missed its take-profit, and the platform failed to match and later account for the broker position. | Running code did not demonstrate that the intended order controls and records were in place. |
| Strategy and portability | The order strategy differed from Ike’s description, while EUR/USD and OANDA were hard-coded across the project. | Both the intended trading rules and the limits of the implementation needed attention. |
| Backtest reproducibility | The earlier historical backtesting did not work well; the later design used stored data and versioned settings. | Fixed evidence and traceable settings were meant to make test results repeatable. |
| Human approval and failure handling | The later design treated broker confirmation as necessary and blocked activity when outcomes were unclear or data disagreed. | Uncertainty was a reason to stop new activity, not to assume the platform and broker remained aligned. |
| Workflow overhead | The atlas-next process grew to 27 workstreams and, at one point, far more documented lines than code. | Ike says resetting the process and changing tools improved planning and reduced rewrites. |
The takeaway is specific to one builder’s experience: an AI agent helped implement a system only after Ike supplied project structure, trading requirements and decisions about how to handle uncertainty. The project’s progress into paper trading did not establish that its strategy had an edge or that it was safe for live funds.
Sources and scope
This retrospective draws on Victor Ike’s first-person accounts: “Three rebuilds: what building a trading platform with AI agents actually took” (September 28, 2026) and “Building my own trading platform, three rebuilds in” (September 25, 2026). Project counts, system behavior and readiness statements above are attributed to Ike; they have not been independently established here.
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.
