Enterprises are using AI to write more code than ever before; anywhere between 25% and 75%, depending on who you ask. This means developers are moving to other parts of the process, where they run into whole new sets of problems.
IBM rolled out its IBM Bob agentic software development platform earlier this year to help developers across the entire software development lifecycle (SDLC), rather than just in single interfaces or isolated tasks.
To build out the platform, IBM Thursday announced a series of updates, including new multi-agent capabilities, parallel tool calling, and built-in cost and use analytics. The company also announced three specialized workflows geared specifically to Java modernization, its IBM i operating system (OS), and its mainframe architecture, IBM Z.
“What makes IBM Bob different is that IBM did not build it as another point coding assistant,” said Michael Kwok, VP of IBM Bob. “The market conversation has moved from ‘which model writes code fastest?’ to ‘which platform helps enterprises deliver software safely, repeatedly, and economically across the full lifecycle?’”
Bob is designed to address that broader problem, he said: understanding complex systems, planning changes, executing work, validating results, and giving leaders visibility into usage, governance, and cost. “IBM Bob supports the work around the code, as much as the code itself,” he said.
Bob, which was made globally available in April, embeds agentic AI across the entire development process: discovery, planning, design, coding, testing, deployment, and operations. It offers different persona-based modes (‘Agent,’ ‘Plan,’ ‘Ask’), reusable playbooks, and enforced standards.
Bob can call tools to perform different tasks and route those tasks between different models, like IBM’s Granite or Anthropic’s Claude, based on cost, performance, and accuracy needs. It can also run several tasks simultaneously, each in its own thread. From a security standpoint, it scans sensitive data, enforces policy in real time, and incorporates red-teaming directly into development workflows.
“Enterprise software work is rarely a single prompt or a single file,” said Kwok, noting that it often requires repository discovery, dependency analysis, testing, security review, documentation, and human approval. “Bob coordinates that work, rather than leaving developers to stitch it together manually,” he said.
Now, rather than running each one separately, Bob can call model-native tools in parallel and run them simultaneously. This means that a task that previously took 30 seconds can now be done in 10 seconds or less, reducing token consumption per task, IBM says. Its context window is also larger (270K tokens compared to 200K in V1).
Additionally, Bob can pull in subagents to perform its exploratory steps. When the agent needs to do a self-contained task, like “figure out how authentication works in this codebase,” it spins up a subagent to read files, perform analysis, and work out patterns. The main agent then receives a summary, and the intermediate steps are thrown away, IBM says. This helps prevent context window bloat.
Parallel tool calling reduces waiting time for work that fans out across searches, file reads, and validation steps, Kwok explained, while subagents keep the main context cleaner by isolating exploratory work and returning concise summaries.
“The point is not that Bob can do more things at the same time; it’s that Bob can coordinate those things in a way that remains understandable, repeatable, and auditable,” he said.
Further, Bob is now equipped with ‘Bobalytics,’ a visibility and cost optimization tool for teams to help them maintain oversight, monitor use, and allocate resources.
“The goal is to help enterprises understand not only how much AI is being used, but if it’s creating meaningful value,” said Kwok.
Bobalytics is designed around multiple views, he explained. For instance, administrators need to see seat usage, consumption, governance controls, and activity visibility, while managers need insight into “team-level patterns,” such as who’s adopting Bob, which workflows are delivering value, and where teams may need support.
This can support important decision-making, Kwok said: Where adoption is high but value is low, teams may need better workflows or training; if a team has cost spikes, leaders need to know where and why, and take action accordingly.
IBM has offered ways to help enterprises modernize across mainframes, Java codebases, and OSes for decades. Now, the company is incorporating that institutional knowledge into three pre-built, customizable workflows for Java modernization, IBM i, and IBM Z. The company says these are “structured, repeatable, auditable, and purpose-built.”
Bob for Java modernization helps teams migrate from Java 8 or earlier to Java 11, 17, 21, or 25, identifying compatibility issues, analyzing dependencies, coordinating code and configuration updates, and performing other important tasks.
For instance, a developer may ask Bob to assess an app for a Java version upgrade. Bob may have to inspect the build system, analyze dependencies, review framework usage, identify compatibility issues, read logs, understand test coverage, and propose an upgrade plan. Now it does those tasks in parallel, while subagents can handle focused investigations “without polluting the main conversation context,” Kwok said. Bob for IBM i features curated skills and agentic workflows optimized for the IBM i OS. This includes refactoring “monolithic” apps into more modular modern structures, creating documentation, producing unit tests, and generating different types of code (COBOL, DDS, CL, RPG) for developers. Further, an ‘IBM i database mode’ allows Bob to emulate an experienced database engineer.
Mainframe environments have been notoriously difficult for AI integrations, and IBM says it is bringing AI-native app modernization to IBM Z for the first time, with COBOL and PL/I modernization and job control language (JCL) analysis.
Bob for IBM Z offers reusable skills; specialized modes that allow it to adapt to different tasks like code refactoring or architectural impact analysis, and the ability to write code, read, files, and execute commands.
For example, a developer may ask: “What impact will this field change have?” and Bob can use Z-specific analysis and metadata to reason across programs, copybooks, JCL, data flows, and subsystem interactions, Kwok noted. A subagent can explore one part of the system, summarize the relevant findings, and return only what the main agent needs to continue planning or executing the change. Java, Z and i are all environments with different runtime assumptions, languages, integration patterns, governance needs, and operational constraints, he said, adding that IBM’s domain expertise is “a key differentiator.”
IBM will eventually broaden into other workflow-specific capabilities, he noted, in areas where “specialized workflows can materially improve real software delivery.”
IBM Bob is not another copilot bolted onto your integrated development environment, said Shashi Bellamkonda, principal research director at Info-Tech Research Group. Rather, it “builds security, testing, and governance into the generation step, so code arrives already checked instead of landing on the reviewers who were the bottleneck.”
Prompt normalization blocks unsafe instructions as they’re written, sensitive data is scanned and secrets detected in real time, and policy enforcement is continuous throughout the code lifecycle, he noted. Bob, rather than a human team, picks models, and built-in and custom models allow developers to move between planning, coding, and review without needing to switch tools. Further, Model Context Protocol (MCP) integration connects Bob to existing toolchains.
Most AI coding tools have typically worked in the same way: Generate code in a coding tool, paste it into an integrated development environment (IDE), then spend time fixing what broke, Bellamkonda pointed out.
Developers end up writing a lot of code and losing hours chasing bugs. Then code hits production, where every line still has to clear security review, testing, and compliance. And while, for example, AWS Kiro requires a spec before any code exists, then tests code against it, AWS Transform goes after the other end, modernizing old code and clearing tech debt in a continuous loop.
“IBM Bob works the same stage but bakes the checks into generation,” Bellamkonda noted.
“The whole industry reached the same conclusion this year: Bolt an accelerator onto an unchanged pipeline, and you move the bottleneck downstream,” he said. “The tools just differ by where they step in.”