RSHB Digest: Agents Are Held Back Not by Models but by Access, Control and ROI

If you boil the October digest from Rosselkhozbank's AI Lab down to a single idea, it goes like this: models are already smart enough, and the problems begin around them. Who checks the result, what rights the agent has, who is responsible for its actions, and how much money all of this actually brings in. The author, Olga Popova, collected the month's trends and added comments from colleagues. This is not a study but the view of practitioners from a bank, yet for those building products on agents it offers useful reference points.

Code is written faster than it is reviewed

According to the authors' observation, a shift is visible at GitHub Universe 2026 and in Anthropic's releases (Mods for Claude Code): the developer is turning from a "code writer" into an architect and task-setter. The agent generates code many times faster, and the bottlenecks become design, security and review.

Imagine an agent producing a large database migration in an hour. Writing it was quick, but a human will have to check it. If the reviewer does not understand the schema, they will sign off on the changes blindly. Hence the paradox from the digest: to use AI safely, engineers must be stronger, not weaker.

The practical takeaway is ours, not the authors': measure the value of agents not by lines of code per day, but by the share of changes that passed review and were not rolled back.

An agent with access is a question of accountability

An agent without access to working systems remains an ambitious chatbot. To deliver measurable value, it needs databases, CRM and ERP. The Russian market is preparing for this: Trukonf showed TrueConf Server MCP, Sber launched the Cowork Bench evaluation environment, and the FinTech Association is preparing unified rules for how neural networks work with financial systems.

OpenAI, as the digest describes it, went further and introduced Dots: agents on GPT-6 Astra that work around the clock on cloud PCs and help in the background. Control is built into the product: Custom Rules, Auto-review and Activity View. Dmitry Zhikharev of RSHB notes that manageability is now designed in from day one, rather than added after an incident.

The topic of payments shows what is at stake. The Bank of Russia is studying scenarios for using agents in payment infrastructure, and the A7 system has already trained AI to check payment documents. Svetlana Bulgakova believes that once an agent can initiate a payment, limits and multi-level confirmation will matter more than the technology itself.

The digest treats cybersecurity separately. Citing Western media, it claims that GPT-5.6 Sol helped find critical vulnerabilities in PHP, that attackers are using agents for attacks for about $5,000, and that OpenAI canceled the release of GPT-6.1 Astra because agents were bypassing protections. Kirill Malkov, the lab's CTO, calls such stories largely speculation and PR ahead of a possible IPO of frontier companies. This cannot be verified from the digest alone. But the conclusion does not change: an agent must live in a strictly controlled environment.

Money and what to do about it

The month's second trend: instead of rankings, speed and cost are becoming more important. Google released Gemini 4 Argon, and researchers at Tsinghua University proposed the Cache-to-Cache method, in which models exchange data through the cache rather than through text. The claimed speedup in joint work is 100–150%. The figure should be treated with caution: the digest does not give the measurement conditions.

The financial picture is sober. By 2026, large Russian businesses may spend 237 billion rubles on neural networks, but according to SberPro and KPMG, most companies know how to launch pilots, not how to calculate the financial result. Shares of a number of Chinese AI companies, according to the same review, fell by more than 40%. The market is beginning to demand figures instead of promises.

What to take into practice:

• Give the agent a separate account with minimal rights and limits on amounts.
• Keep an action log and show it to a human. Confirm irreversible operations manually.
• Calculate ROI including infrastructure, not by the number of licenses and tokens.
• Invest in review skills: this is the new bottleneck.

The limitation of the material is simple: it is an author's review by a corporate lab with expert opinions. Many facts are given secondhand, without primary sources, and some claims about models and incidents require independent verification. As a map of trends it is useful; as a basis for decisions it is not. The digest's main question remains open: where do you draw the line of autonomy? Would you let an agent into production code, into corporate systems, or near payments?