Pyotr Tripolsky on Habr showed an AI agent for analyzing news and Bitcoin signals

Peter Tripol'skii published on Habr a scheme for an AI agent for news analysis of the crypto market. In the article, he breaks down why popular GitHub solutions for trading signals break on real-world events, and shows a more practical scheme: the agent first searches for fresh market triggers, then outputs a short signal BUY, SELL, or WAIT.

Where Analogues Break

The author starts with criticism of two popular approaches. The first is swarms of agents that debate each other based on raw indicators like RSI, Stoch RSI and other metrics. On paper this looks convincing: one agent proves growth, another proves decline, and the user chooses the number of debate rounds. But in reality, writes Tripol'skii, the model often clings to the first signal it encounters and builds an explanation around it, ignoring conflicting data. As a result, the system looks more like a beautiful simulation of analysis than a real decision-making tool.

The second approach seems more solid: several agents divide fundamental topics among themselves — ETF flows, on-chain reserves, hash rate, actions of large holders, macroeconomics and price. The problem is that such analysis is too static. It may be useful for the general picture over a month, but responds poorly to sudden market shocks, when price is moved not by the balance of indicators, but by one sharp event at a specific hour. It is here, in the author's view, that the market manages to go in a completely different direction.

New Search Logic

Instead of a debate between abstract agents, the author proposes a scheme with the "reasoning + action" pattern. The point is that the model should not close itself within a pre-loaded set of metrics. Between reasoning steps, it needs to be given the opportunity to search for facts anew on the internet and clarify exactly what happened in the last few hours. So the agent is not discussing the market in general, but responding to the specific context of the date and moment. This is an attempt to make the query to the world adaptive, not one-time.

"Sharp events outweigh lagging analysis".

In the published code, the web-search agent is given strict rules: don't peek into the future, don't use materials without a clear date, and don't copy one opinion from an article as a ready-made conclusion. In essence, the agent works like a fast editor of market summaries, who should collect several confirmations before issuing a signal. For this, it is offered to make several search queries and cross-check the picture across independent publications.

• Searches for breaking news on the asset over the past 4–12 hours
• Checks regulatory actions, exchange hacks and withdrawal halts
• Tracks macro-surprises like Fed decisions or CPI publication
• Looks at abnormal volumes and reasons for sharp price movements
• If the picture is contradictory, chooses WAIT instead of a forced forecast

Separately, the author formalizes the output format. The model is obliged to return not a long review, but one of three actions and a short explanation of which events led to it. This brings the system closer not to an analytical essay, but to an operational tool that can be connected to a backtest or trading loop. This format is especially convenient where the signal then needs to be automatically checked on history or passed on to trading logic.

Market Testing

To demonstrate, Tripol'skii ran the agent on Bitcoin news for April 2026. Over a long horizon, the general fundamental analysis looked bearish: high rates, weak volumes, pressure from miners and ambiguous institutional flows pushed the model towards a SELL signal. But then the article shows why such a framework alone is not enough. The monthly picture and the decision for the coming hours turn out to be different tasks. It is this gap that the author tries to close with a news agent.

On April 5, 2026, the agent saw a rise in geopolitical tensions following a US ultimatum to Iran, but did not find enough clear and unambiguous catalyst for entry. The result was WAIT. Already on April 8, 2026, after an announcement of a two-week ceasefire with Iran, the system recorded a sharp jump in Bitcoin to $72,000, liquidation of approximately $425 million in short positions and a surge in volume — and issued BUY. And on April 9, 2026, the signal changed again to WAIT: the bullish impulse collided with pressure from miners, options and new risks, so the model chose caution.

This fragment is the main idea of the entire article. The author does not promise a "smart oracle" that always guesses the market. On the contrary, he shows a more disciplined mechanic: first identify the sharp event, then verify confirmations, and only then decide if there is even a basis for a trade. For AI agents in finance this is more important than beautiful reasoning about fundamentals, because the error here arises not from a lack of data, but from the wrong order of working with it.

What This Means

The article on Habr offers not a ready-made trading grail, but a useful scheme for news AI agents. The key idea is simple: in a rapidly changing market, the model should be able to rediscover context by date, acknowledge uncertainty and more often choose WAIT if there are not enough fresh facts. Such an approach may be useful not only in crypto, but in any systems where the decision depends on news from the last hours and minutes.