Why better coding agents do not guarantee better returns

Widespread AI adoption leaves the economic question open

A coding agent can become more capable while remaining hard to justify as an investment. For engineering leaders, the business question is whether it improves software delivery enough to cover the extra resources it consumes. Wider adoption and better task performance make that question more urgent, but neither settles it.

Consulting firm McKinsey describes this gap in its McKinsey Technology Trends Outlook 2026. As a consultancy advising businesses on technology, McKinsey has a commercial interest in companies seeking help with AI adoption and integration. Its report says 89% of companies regularly use AI and a majority are at least testing agents. Yet financial gains remain less common: “We find that only 37% of companies attribute any positive EBIT impact from their AI programs, let alone from newer agentic deployments,” McKinsey writes. EBIT means earnings before interest and taxes, a measure of operating profitability.

These figures cover AI programs broadly, which limits what they tell engineering leaders about coding agents. They establish that deploying AI and earning a financial return are separate achievements. A coding-agent investment therefore needs its own assessment of delivery improvements and the costs of achieving them.

Autonomous-task benchmarks measure technical progress

That financial assessment starts with understanding what agents can do. McKinsey credits Anthropic and OpenAI with improvements in tool use and computer interaction that let agents carry out more development actions. It also identifies LangGraph, associated with LangChain, as providing workflow coordination and “human-in-the-loop” controls, through which people can review or guide an agent’s work. Anthropic and OpenAI sell AI models, while LangChain supplies tools for building AI applications, so all three have a commercial interest in wider agent use.

These capabilities matter because software development involves connected tasks. McKinsey describes potential agent roles in analyzing requirements, generating code, conducting tests, finding bugs, and handling parts of deployment. An agent that can move through several of these activities has a broader practical role than a tool answering an isolated request. Measuring that ability requires a clear definition of autonomous task completion.

Model Evaluation and Threat Research, or METR, is a nonprofit that benchmarks independent task completion. Its time-horizon measure expresses task size as the average time a human would need to finish the work. METR reports that horizon at a specified success rate, linking task size to the probability that a model will complete it successfully.

For Claude Opus 4.6, METR’s cited software-engineering horizon is approximately twelve hours at a 50% success threshold. The model is therefore expected to complete half the cases involving tasks that would take a human about twelve hours on average. This describes the amount of human work represented by the task. It does not describe the model running independently for that duration.

The distinction matters because McKinsey appears to have interpreted the horizon as elapsed AI runtime. METR researcher Sydney Von Arx explained the measure in an interview with MIT Technology Review and said she initially questioned whether it was the right one. METR nevertheless observed longer horizons among leading models and an accelerating improvement trend. “I can speculate all I want about whether this makes sense or not, but the trend is there,” she said.

The trend shows increasing autonomous capability, but its business value depends on how companies use it. McKinsey describes agentic AI, systems that take actions across multiple steps, as early in its maturity and expects larger benefits as the technology develops and becomes more deeply integrated into enterprises. Engineering leaders have reason to keep evaluating those changes, while separating capabilities available now from improvements expected later.

Agent workflows add computing demand and infrastructure costs

Using those capabilities consumes resources at each step of an agent’s work. McKinsey calculates that a single agent workflow can require five to 30 times the computing resources of a typical chatbot query. An agent workflow may contain many operations, whereas the chatbot comparison covers one query. The multiplier therefore describes the demands of different units of work, and its relevance to a deployment depends on what the workflow contains.

Those demands follow from how agents operate. Repeated inference loops require a model to process information and decide its next action; tool calls let it act through other software; retrieval steps bring in information needed for the task. Coordination across systems adds further work. As these activities recur, computing demand grows beyond the resources needed to process the initial instruction and response.

The surrounding technology adds to the cost. McKinsey says companies need to upgrade their existing technology stacks, the software and infrastructure on which their systems run, to support agents reliably. A deployment budget therefore has to cover both the environment needed to operate the agent and the computing resources consumed each time it works.

These requirements are already putting pressure on budgets. In a survey McKinsey describes as recent, 93% of participants said additional resource demand had caused them to exceed their AI budgets. The finding applies to those survey participants, but it gives leaders a concrete reason to assess operating costs alongside expected benefits. Higher spending can be worthwhile when improvements in completed development work are large enough to pay for it.

Development outcomes vary sharply across companies and developers

The potential improvement is large enough to warrant serious evaluation. McKinsey estimates that successful agent deployment in software development could create almost a trillion dollars in additional value for companies. Realizing that potential depends on whether agents improve how development work reaches completion. Company-level results show how hard that can be.

In a study reported by McKinsey, only a quarter of companies using software-development agents significantly accelerated their development cycles. McKinsey defines significant progress through two requirements together: productivity must at least double, and the improvement must cover more than a quarter of teams. A strong result in a small group can therefore fall below the threshold. Companies outside that threshold may still obtain smaller benefits worth paying for, but the measure tests whether a gain becomes substantial across an organization.

Meeting that organizational threshold is harder when individual gains are uneven. In research reported by McKinsey, around 80% of developers using AI tools achieved productivity improvements of about 3%, while the top 20% achieved improvements of 55%. These findings cover AI tools broadly, so they offer a wider view of developer outcomes than coding-agent results alone. They also show why a deployment budget needs to account for how benefits are distributed: the strongest individual result can differ greatly from what most developers experience.

Company-level results can also worsen after adoption. McKinsey reports that productivity declined in 30% of companies after their development teams began using coding agents. That sequence leaves the cause of the decline open, but it shows that increased use can coexist with a worse organizational result. Leaders therefore need to distinguish individual productivity, development-cycle speed, and company productivity before using any one measure to justify spending across teams.

Even when measures improve together, their rates of improvement can tell different stories. In one study reported by McKinsey, AI tools were associated with a 180% increase in programming activity but a 30% increase in published releases. Releases measure output further along the development process, after more of the work needed to produce software is complete. Measuring programming activity alone would therefore give a more optimistic account of the change.

The release count also needs to connect to work the company values, because more releases do not establish their quality or economic contribution. Leaders can start an evaluation with the development activity an agent is meant to improve, then examine what happens to completed software. A faster coding step has a different economic value depending on whether testing and deployment also improve. This connects the measured benefit to the full process that must bear the agent’s computing and infrastructure costs.

Developer distrust creates a separate adoption challenge

Delivery outcomes describe what happens when teams use AI tools; developer confidence describes another condition for their use. McKinsey reports that around 46% of developers worldwide actively distrust AI-tool accuracy, a third trust the tools in principle, and 3% have great confidence in their results. These figures describe different levels of confidence and should be read individually, because they do not form a complete breakdown of developer responses.

Alongside those confidence figures, McKinsey characterizes some programming as proceeding by “feel.” That description concerns working practices, while the survey figures concern confidence in tool accuracy. Confidence remains relevant to adoption, but the figures alone do not establish how skepticism contributes to productivity variation or budget pressure. Leaders need to assess developer trust separately from an agent’s technical ability and its measured delivery results.

Investment and hiring reflect expectations of growth

Despite these mixed operating results, financial activity around agent-based software development has grown sharply. McKinsey reports that investment in providers rose from a negligible level in 2023 to around $5 billion in 2025, then exceeded $61 billion during the first half of 2026. Related job postings increased 221% between 2024 and 2025. Investment and hiring indicate expectations of future demand and a willingness to commit resources to serving it.

The investment total also needs to be understood through its composition. McKinsey attributes most of the first-half surge to SpaceX’s acquisition of Cursor for $60 billion in stock. Because that transaction accounts for most of the reported amount, the increase is highly concentrated. The headline total captures a large acquisition alongside other investment in the category, rather than evenly distributed funding growth across providers.

Key takeaways for leaders

• Assess coding-agent returns separately: McKinsey reports that 89% of companies regularly use AI, while 37% attribute positive EBIT impact to their AI programs. Engineering leaders need a deployment-specific assessment linking delivery improvements to total costs.
• Read autonomy benchmarks carefully: METR’s twelve-hour horizon for Claude Opus 4.6 represents human task time at a 50% success rate. Engineering evaluators need to test completion rates on their own development tasks before budgeting for autonomous work.
• Budget for the full workflow: McKinsey estimates that an agent workflow can consume five to 30 times the computing resources of a typical chatbot query. Technology owners need to include repeated model calls, tool use, retrieval, and infrastructure upgrades in deployment budgets.
• Measure completed software: Only a quarter of companies using development agents met McKinsey’s threshold for significant acceleration, while 30% reported productivity declines after adoption. Engineering managers need to track releases, quality, and development-cycle speed alongside the cost of achieving those outcomes.
• Evaluate developer trust separately: Around 46% of developers actively distrust AI-tool accuracy, according to McKinsey. Engineering managers need to assess confidence alongside delivery results; the survey alone establishes no causal link between distrust and productivity.
• Account for investment concentration: SpaceX’s reported $60 billion stock acquisition of Cursor accounts for most of the investment surge cited for early 2026. Procurement and strategy teams need to assess provider maturity and customer outcomes separately from headline funding totals.