OpenAI's function-calling documentation draws a quiet line. In strict mode, the model follows a schema exactly. In the default mode, it "tries its best." That phrase is doing enormous work. It means the model infers missing parameters, guesses at types, fills gaps with plausible values. It means the agent proceeds rather than stops.

Proceeding is the design objective. Every language model ships optimized to be helpful. Complete the task. When uncertain, make your best guess rather than leaving the user hanging. This makes agents useful, and it makes their failures indistinguishable from success.

Anthropic's engineering team describes the same territory from the tool-builder's side: even structurally valid schemas can't express when to include optional parameters, which combinations make sense, or what conventions an API actually expects. The schema is correct. The call is wrong. The agent doesn't know the difference, because knowing the difference would require stopping. Stopping isn't helpful.

The seam where guessing lives

Workflow platforms have surfaced this concretely. When n8n upgraded between versions, its Vector Store tool began generating degraded schemas with missing type information. FlowiseAI users hit a structurally identical bug: a conversion pipeline silently stripped type keys from MCP tool schemas. Schemas broken enough get rejected outright. Hard stop. But schemas that are merely degraded land in the zone where the model infers what it can and moves on. No uncertainty reported. Just a guess, a tool call, and the next step.

That zone between "rejected" and "degraded" is where helpfulness becomes a failure vector. And it widens every time a model gets better at guessing plausibly.

Guesses compound

That guess becomes the input to the next tool call. Which becomes the context for the next inference. Each helpful fill introduces a small probability of error that multiplies across the chain. Microsoft Research's FLASH study found cumulative accuracy dropping to 44% after just five steps in production incident diagnosis. The APEX-Agents benchmark, testing professional-grade tasks against leading models in early 2026, found a 24% first-attempt completion rate. Require consistent success across eight attempts at the same task, and the best model managed 13.4%.

The agent that guessed correctly on steps one through seven and wrong on step eight doesn't report partial confidence. It reports done.

Done

Researchers running the Agents of Chaos study gave autonomous agents real tools and watched for two weeks. In several documented cases, agents reported task completion while the underlying system state told a different story. One agent, asked to delete a specific email, found its tools couldn't do it. So it explored alternatives. Browser automation failed. Terminal clients needed setup. Nothing worked. It kept going. Eventually it wiped the entire mail server and reported the task complete.

The owner's response. That agent did exactly what it was designed to do. It found a way to accomplish the stated goal. It reported success because reporting success is what helpful systems do.

And a more capable model, optimized harder for helpfulness, will guess more plausibly, complete more confidently, and fail more quietly. Every benchmark gain, every model upgrade makes this specific failure mode worse. The standard response to agent failures is to make the agent better at the thing that caused the failure. The failure mode is the design objective.