From Code to Delivery: Controlling the Cost of AI Coding Agents
At ALIGN Chicago 2026, IBM's Will Hawkins argued that coding agents are here to stay but need oversight, predictable pricing and analytics that tie spend to production results across the software lifecycle.
By The AI Enterprise Magazine Team3 min read
Key takeaways
- Skip vibe coding. Experienced developers must review what agents produce.
- Agents are here to stay. Hawkins said productivity gains are too large to go back.
- Costs are hard to forecast. Model choice and token types make budgets unpredictable.
- Route by task. Simple work like documentation doesn't need the largest model.
- Measure what ships. Track spend per team against commits and accepted code.
- Cover the whole lifecycle. Planning, testing, deployment and maintenance matter too.
Almost everyone in the room at the ALIGN AI Executive Summit Chicago 2026 raised a hand when asked if they had used a coding agent. Will Hawkins, PhD, a data and AI technology leader at IBM, wanted to talk about what comes next: getting enterprise value from those agents while controlling cost and keeping quality high. He spoke on July 15, in a session sponsored by IBM titled From Code to Delivery: Modernizing the SDLC With Agentic AI.
Why not just vibe code?
Hawkins was blunt: "Vibe coding is a terrible idea." Code generated without review can carry security flaws and inefficiencies and often never makes it into production. "You should absolutely not trust a large language model to blindly build your core business applications," he said. Instead, he argued for agentic engineering, with experienced developers overseeing what agents produce.
He was equally clear that coding agents are here to stay. The tools aren't perfect and need oversight, and they produce better results in the hands of strong developers. But he said the productivity gains are large enough that software development won't go back to working without them.
Why are coding agent costs so hard to predict?
Hawkins said most enterprises have no clear idea what they will spend on coding agents next month or next year, or what productivity they will get in return. Lines of code are a poor proxy for productivity, he added, and may better measure inefficiency.
The pricing itself is complex. Many tools let developers pick from models whose prices vary widely. Costs are split across input tokens, cached input tokens and output tokens, and each depends on how much context developers send, such as an entire program for a security review. Given the choice, he said, engineers tend to pick the most expensive frontier model even for routine work like documentation. For executives responsible for engineering budgets, that makes planning nearly impossible.
How does IBM approach the problem with Bob?
Hawkins described several ways IBM's coding agent, Bob, addresses cost:
- Flat pricing. He said IBM charges the same per-million-token rate whether tokens are input, cached or output, making spend easier to forecast.
- Automatic model routing. Users don't choose the model. Bob classifies the task and its context, then routes it to the most cost-efficient model that can produce a quality result. Writing documentation for an old Java application doesn't need the largest model available.
- Usage analytics. A reporting layer shows spend per developer and per team, supporting chargebacks, alongside commits that reach production, how much suggested code developers accept and how sprint times change through integrations with tools such as Jira and Azure DevOps.
What does a full-lifecycle code agent look like?
Most competing tools focus on writing code, Hawkins said, but that is only one stage. Bob is designed to help plan an application, design it, write and test code, deploy through MCP connections and pull requests into GitHub, and support maintenance. Because each stage often has its own team and tools, connecting to all of them through MCP matters. Bob runs as a VS Code-based environment and a command-line tool.
IBM also offers add-ons for modernization, including one for the mainframe. It scans applications, extracts metadata and maps dependencies so teams can modernize legacy COBOL written by developers who retired long ago. Similar packages cover IBM i and Java modernization. Hawkins said on-premises versions are coming, allowing organizations to shift coding agent spend from operating to capital expense instead of paying by the token.
What results has IBM seen internally?
Hawkins said about 100,000 IBM employees now use Bob, including people in sales and finance. He said 95 percent of those given access use it, 65 percent use it daily, and IBM has measured an average productivity improvement of 45 percent.
Where to hear more in 2027
Technology leaders will keep comparing notes on coding agents and AI costs at the 2027 ALIGN AI Executive Summits, invitation-only gatherings in eight U.S. cities: Atlanta, New York City, Dallas, Chicago, San Francisco, Boston, DC and Los Angeles. Registration is coming soon. See the ALIGN AI Executive Summits page.
Frequently asked questions
Why is vibe coding risky for enterprises?
Hawkins said unreviewed AI-generated code can include security issues and inefficiencies and often never reaches production.
Why are AI coding agent costs hard to predict?
Prices vary by model and by input, cached input and output tokens, and usage depends on how much context developers send.
How can teams control coding agent costs?
Hawkins described routing tasks to the most cost-efficient capable model and tracking spend per developer and team against results.
This article is based on remarks by Will Hawkins, PhD, Data and AI Technology Leader, IBM at ALIGN AI Executive Summit Chicago 2026 on July 15, 2026. More coverage: ALIGN AI Executive Summits.
Filed under: Agentic AI, AI Strategy & ROI, AI Governance & Trust
