AI coding agents such as Claude Code, Codex and Kimi Code are pushing software development from human-written code with tool assistance toward a workflow in which humans define goals and agents execute for extended periods.

The metric is shifting away from tokens

During an InfoQ livestream at the 2026 World Artificial Intelligence Conference, Sirius contributor Teng Yu and Moonshot AI developer relations lead Tang Feihu discussed how AI coding is changing engineering practice, management and career growth. Teng tends to buy the strongest models and top-tier plans because he values problem-solving capacity over fine-grained price comparison. Tang tries a wider range of tools, including Kimi Code, Claude Code, Codex and vertical products, then chooses model-and-harness combinations by task.

Token usage is becoming a weak proxy for productivity. Earlier “vibe coding” often meant one prompt and one answer. Today, a developer may give an agent a target and let it call tools, edit files, debug and continue for an hour before human review. That process can waste reasoning steps and tool calls, but it may still complete the job. For companies, the more relevant questions are whether the task was finished, how much human intervention was required, what failure would cost and whether the output created business value.

“Everyone can build” still has limits

Both speakers agreed that AI lowers the barrier for non-programmers, but not without boundaries. Teng argued that agents can often build applications centered on interfaces, workflows and common business logic if requirements are clearly described. However, when a task requires a new architecture, unfamiliar logic or a direction not well represented in existing code, models can loop on the wrong approach. He estimated that 70%-80% of ordinary application requirements may be achievable through vibe coding, while critical parts still need human judgment.

Tang was more optimistic. He recalled trying to build a 3D boat-parking game for a SIGGRAPH-related activity: earlier models failed to produce a satisfying result after 48 hours, while the latest Kimi K3 generated a playable version from a single instruction. The point is not that expertise disappears, but that model progress can turn tasks once reserved for developers into natural-language requests. Still, generating code is not the same as being an engineer. Humans must decide why a product exists, whether a target is sensible and whether the final delivery satisfies real needs.

Models are engines; harnesses make them usable

In AI coding systems, the model supplies core capability, while the harness provides tools, context, execution environments and workflows. A harness is the surrounding system that connects a model to real tasks. The same model can behave very differently when placed inside different harnesses.

Teng warned that prompts and workflows can become obsolete quickly as models improve, so over-investing in a “perfect” harness too early may be inefficient. He puts more weight on product direction, team communication and open-source collaboration. Tang compared models to aircraft engines and harnesses to the airframe: the engine is essential, but it cannot fly alone. Model companies build harness teams because they capture user behavior and feed real needs back into training teams. Vertical harnesses can also integrate MCP, Skills and business tools to raise success rates in specific scenarios.

Key variables now include:

• model capability and reasoning intensity;
• tool access, context handling and workflow design;
• internal benchmarks and security rules;
• user experience, token cost and task success rate.

Responsibility and ROI favor experienced engineers

The hardest question is not whether AI can write code, but who is responsible for the result. Teng argued that even strong models rarely decide that a direction should simply be stopped, and they cannot bear the consequences of production failure. Tang said models already show forms of judgment, such as evaluating writing quality or refusing unsafe actions. These views can coexist: a model may generate a judgment, but a company still needs a human accountable for it.

That matters in production. Even if AI can write, debug and test code, post-deployment incidents, risk trade-offs and accountability cannot be handed entirely to a model. Companies still need to evaluate whether agents call the right tools at the right time, follow internal rules and actually improve task success rates.

ROI is therefore not about giving everyone the same model budget. For complex tasks, repeatedly using an unsuitable cheaper model may cost more than using a stronger model from the start. Teng went further: the best models should often go first to the most experienced engineers, because they know which tasks matter, how to decompose them, when to correct the agent and when to stop.

The likely direction is flatter software organizations and more leverage for small expert teams and open-source groups. The old path in which junior developers grew mainly by repetitive coding will be compressed. The durable engineering skill will be less about typing every line and more about defining problems, validating outcomes, taking responsibility and turning agent execution into reliable delivery.