Developer argues AI agent tools remain 'vapourware' amid usability and cost concerns
A recent opinion piece highlights the gap between marketing promises and practical application, citing login failures and excessive compute costs as evidence of poor quality assurance in the sector.
Software engineer Kira Howe has published an opinion piece titled "Is it all just vapourware?" on her personal website, arguing that current AI agent software is largely ineffective due to significant usability friction and poor quality assurance. The article details a negative user experience with the AI development platform ONA, citing initial login failures on the desktop version and excessive consumption of paid "compute units" for minimal progress when attempting to integrate with Linear.
Howe contends that while the concept of automated, end-to-end agentic development is appealing, current tools add complexity rather than efficiency, creating more work for developers through debugging integrations and managing tokens. The piece reflects a growing skepticism among developers regarding the practical application of "agentic" AI in software development workflows, highlighting a broader industry trend where AI software is built by engineers with unlimited token access but lacks systematic review or QA processes for usability testing.
The specific platform criticised is ONA, which Howe attempted to use for automating the refinement of vague requests into implementable features, design, testing, and deployment. Howe reports spending nearly $20 worth of "ONA compute units" while the system struggled to retrieve tasks from Linear, describing the process as "thrashing." This experience underscores the disconnect between the promise of streamlined automation and the reality of fragmented, error-prone user interfaces.
The article highlights that the cost of writing code has reportedly collapsed due to LLM inference capabilities, leading to an abundance of "vibe coded" startups, though execution often lacks regard for user experience. Howe suggests that if agentic development were truly effective, vendors would be the primary beneficiaries, yet their own products often fail basic usability tests, such as the authentication hurdles encountered on ONA’s desktop application.
Howe notes that while she is willing to pay for developer tooling that removes distractions from core coding tasks, current offerings frequently result in more annoying work, including debugging integrations and managing auth tokens. The criticism points to an industry where marketing claims outpace product maturity, leaving developers to manage additional overhead rather than gaining the promised efficiency gains from automated software development lifecycles.
