What Bob can actually do and how it works

Bob is not a single model, but a system of specialized agents, each with its own role — planner, coder, tester, deployer, and operations manager. Above them sits the Orchestrator Agent, which coordinates their work and inserts mandatory human approval points (so-called checkpoints). The development team can thus set exactly when a human must approve the next step — for example, before deploying to production.

Bob's architecture supports multi-model routing: based on the type of task, it automatically selects between IBM Granite, Anthropic Claude, Mistral, or lightweight local models. This allows cost scaling — simpler tasks use cheaper models, while complex reasoning tasks use more expensive ones. Integration via the Model Context Protocol (MCP) enables connecting external tools, APIs, or on-premise servers without losing context.

Developers work with Bob through an IDE plugin, CLI interface, or Bob Shell. Across all interfaces runs a unified governance layer that enforces company policies, logs every action, and displays approval prompts. For regulated industries such as banking or healthcare, this consistency is crucial.

Pricing: Bobcoins as the new currency of development

IBM has chosen an unusual billing model — a Bobcoin credit system, where one Bobcoin equals $0.50. The offering stands on four tiers:

• 30-day free trial — 40 Bobcoins
• Pro — $20/month, 40 Bobcoins
• Pro+ — $60/month, 160 Bobcoins
• Ultra — $200/month, 500 Bobcoins

Enterprise customers receive centralized administration and overview dashboards. IBM also offers the Premium Package for Z, a special version for mainframe clients — an area where IBM traditionally dominates and where competition is virtually nonexistent. The credit model may complicate budgeting in companies that already have established chargeback policies, but it also provides predictable variable costs directly tied to actual agent usage.

Security: PromptArmor uncovered a vulnerability

In January 2026, researchers from PromptArmor revealed a series of prompt-injection attacks that could trick Bob into executing malicious shell commands — but only if the user had the "always allow" setting enabled. The attack bypassed several built-in defenses by exploiting gaps in process substitution parsing. IBM acknowledged the finding and promised fixes before the public launch.

This incident underscores a broader truth about agentic AI: the more autonomous an agent is, the greater the identity and delegation risk it brings. IBM has since emphasized continuous red-teaming, prompt normalization, and policy enforcement as the main pillars of security. For any company considering Bob, it is crucial to verify which mitigations have actually been deployed since January.

Competition and market context

According to MarketsandMarkets, the AI coding tools segment is expected to reach $12.6 billion by 2028. Already operating in this field are GitHub Copilot, Amazon Q Developer, Google Antigravity, Cursor, and a number of startups. IBM differentiates itself in three ways:

• Governance and compliance — a built-in layer that is mandatory, not optional
• Multi-model routing — flexibility in AI model selection based on the task
• On-premise deployment — the ability to run in one's own data center, which is essential for regulated sectors and government administration

At the same time, MIT Sloan researchers warn that agentic AI requires new governance frameworks. Identity management must reflect that agents act continuously — as opposed to static human permissions. Companies need a real-time audit trail, the ability to immediately revoke access, and a policy engine that understands chained operations. IBM addresses this natively, but the responsibility for proper configuration remains with the customer.

What this means for Czech developers and companies

Bob is available globally, meaning for Czech companies as well. IBM has a strong customer base in the Czech Republic — financial institutions, telecom operators, and government administration run on IBM infrastructure. This is precisely where Bob could find adoption most easily, because companies know the watsonx ecosystem and have established relationships with IBM.

However, Czech specifics also bring challenges. Bob does not communicate in Czech and does not generate localized code for the Czech environment — documentation, comments, and prompt engineering will need to be conducted in English. For smaller Czech companies and startups, it may be more appealing to stick with cheaper alternatives like Cursor or augmented Claude Code, which offer similar functionality without enterprise overhead.

For large corporations in the EU, the question of data sovereignty is also key. Bob supports on-premise operation, thereby avoiding risks associated with transferring proprietary code to the public cloud. This is a crucial argument for companies preparing for the full effectiveness of the EU AI Act.

How Bob changes developer skills

The rise of agentic coding redraws the map of required skills. Developers will no longer primarily write code line by line, but will learn:

• Prompt engineering — the art of formulating tasks so that the agent understands the context
• Policy writing — defining rules that the agent must not violate
• Multi-model tuning — selecting the right model for a specific task
• Security oversight — continuous monitoring of the agent's actions and audit logs

This is not about replacing developers, but about a shift from writing code to managing development. Juniors can get involved in more complex projects faster, seniors gain time for architecture and design. IBM itself emphasizes that Bob is an "AI development partner," not a programmer replacement — human checkpoints are an integral part of the architecture.