After the Graveyard: Agentic AI Survival

In The AI Agent Graveyard we made an uncomfortable argument: roughly nine out of ten healthcare AI initiatives never reach sustained production, and almost none of them die because the model was wrong. They die because governance, auditability, and deployment discipline were treated as post-pilot concerns.

That article was a diagnosis. This one is a forecast — and a deliberately adversarial one. Because the honest question a payer CIO or health system CTO should be asking in 2026 is not "is agentic AI real?" It is: "if the last generation of AI pilots died in the integration layer, why would autonomous multi-step agents — which depend on that layer far more heavily — do any better?"

It is a good question. It deserves the strongest version of both answers.

• ▸Reliability and capability diverge as task length grows. Benchmarks report single-attempt success; production demands repeated success across long horizons. Recent research shows these two properties separate systematically with task duration.[3]
• ▸The failure surface moved. Tool-using agents are brittle not because they reason poorly but because tool-use errors produce irreversible side effects that backtracking cannot repair.[2]
• ▸The fix is architectural, not model-selection. Decomposition into massively-verified atomic steps has been demonstrated to carry an LLM process across more than a million sequential steps with zero errors — by engineering around the error rate rather than waiting for it to disappear.[1]

Agentic AI is not going to succeed because models get better. It is going to succeed in the organizations that put a deterministic, auditable transaction layer underneath the probabilistic one. Everyone else builds a bigger graveyard.

Most vendor content strawmans the skeptics. We will not. Here is the bear case as a serious engineer would make it.

1. Error compounds multiplicatively, not additively

An agent that performs a ten-step workflow at 97% per-step accuracy completes the full workflow correctly about 74% of the time. At thirty steps, 40%. At one hundred steps, 5%. This is not a rhetorical flourish — it is the central empirical finding in long-horizon agent research. Cognizant AI Lab's MAKER work opens by noting that in canonical sequential benchmarks, LLM processes "inevitably become derailed after at most a few hundred steps" because of a persistent, irreducible per-step error rate.[1]

Healthcare workflows are long. A prior authorization intake-to-determination path touches eligibility, benefit configuration, clinical criteria, medical policy, documentation sufficiency, and notification — each with branch logic. Chain them naively and the math is brutal.

2. Tool errors are often irreversible

Mathematical reasoning tolerates a wrong step because the agent can backtrack. Tool-calling does not. AgentProcessBench frames this precisely: tool-use failures "frequently induce irreversible side effects," which is why step-level process quality — not final-answer accuracy — is the metric that matters for tool-using agents.[2] An agent that posts an adjustment to a claim, writes a status to a provider portal, or sends a member notification has changed the world. There is no undo in an X12 278 response that already left the building.

3. Non-determinism collides with audit obligation

A regulated decision must be explainable, reproducible, and attributable. An agent that reaches the same conclusion by a different path on each invocation satisfies none of those by default. ONC's HTI-1 final rule made this concrete for certified health IT: predictive decision support interventions must expose standardized "source attributes" — including training data description, intended use, validation, and known risks — so that users can evaluate them.[6] "The agent decided" is not a source attribute.

4. The reputational overhang is real

Clinicians are not neutral observers. In the AMA's national prior authorization survey, 61% of physicians said they are concerned that health plans' use of AI is increasing prior authorization denials, and 93% reported that prior authorization delays care.[7] Any agentic deployment that touches utilization management inherits that distrust on day one. Deploy autonomy into that environment without visible human control points and you will not get a technical failure — you will get a political one.

5. Evaluation is immature

Most teams still evaluate agents the way they evaluated models: a benchmark score, a demo, a spot check. Reliability science for long-horizon agents argues that benchmarks reporting only pass@1 on short atomic tasks are structurally blind to production behavior.[3] If your eval harness cannot answer "what is the failure rate of step 7 across 5,000 runs," you do not have an eval harness.

That is the bear case. It is largely correct — as a description of how agents are currently built.

The bull case is not "the models got smarter." It is that the engineering discipline around the models matured in four specific ways.

Decomposition beats scale

MAKER's result is the most important data point in this debate: by decomposing a task into maximally atomic subtasks and applying multi-agent voting and error correction at each step, the authors solved a task requiring over one million sequential LLM steps with zero errors.[1] The per-step error rate did not go to zero. The architecture drove aggregate error to zero. This is the same insight that made unreliable commodity hardware into reliable distributed systems, and unreliable networks into reliable TCP.

Structured output and tool contracts

Constrained decoding and typed tool schemas turned the model's output from prose-to-be-parsed into a validated contract. An agent that cannot emit a malformed X12 segment or an out-of-vocabulary CPT code cannot make that class of error at all. A large share of "hallucination" in production systems was never a reasoning failure — it was an unvalidated interface.

Deterministic orchestration around probabilistic steps

The winning topology is not "an agent that does everything." It is a deterministic state machine that calls a model for the specific judgments only a model can make, with every state transition, input, and output logged. The probabilistic component becomes a bounded function call inside an auditable process — the inverse of the 2024 pattern.

Process-level observability

Step-level evaluation, trace capture, and reliability measurement across repeated runs are now standard practice rather than research artifacts.[2][3] You can finally answer "which step is failing and how often" — which is the prerequisite for ever letting an agent operate unattended.

Both cases are true. Agents built as a prompt plus tool access will keep dying at month eighteen. Agents built as a deterministic transaction layer with bounded probabilistic steps are already running production workloads.

The distinction is not subtle, and it is entirely visible in the design documents.

This is the pattern BCP implements. It is deliberately boring where it needs to be.

1. Supervisor / worker topology

A deterministic supervisor decomposes the workflow into atomic units, dispatches each to a narrowly-scoped worker with exactly the tools that unit requires, and validates the result before advancing state. Workers hold no ambient authority. Least privilege applies to agents exactly as it applies to service accounts.

2. Typed tool contracts

Every tool is a versioned schema with explicit pre- and post-conditions. The orchestrator rejects any call that fails validation before it reaches a downstream system. This is where FHIR resource profiles, X12 implementation guides, and internal domain models do real work — they are not just data formats, they are the agent's guardrails.

3. A deterministic transaction layer underneath

Every write passes through a transactional boundary that provides idempotency keys, staged commits, reversibility, and provenance on every mutation. This is the same layer we described in No Data at Rest — PHI is processed in volatile memory, never persisted at rest in the agent tier, and observability captures summaries and identifiers rather than payloads.

The single highest-leverage decision in an agentic program is putting the transaction layer in before the agents, not after.

4. Bounded autonomy with explicit gates

Autonomy is granted per action class, not per agent, and it is earned by measured reliability.

Most healthcare organizations should be running L1–L2 in production today and piloting L3 in narrow, reversible, high-volume administrative workflows. Anyone selling L4 in clinical decision paths is selling next year's compliance finding.

5. Evaluation as infrastructure

Golden datasets per step, adversarial cases, replay of production traces, reliability tracked as a distribution rather than a score, and regression gates in CI. Model upgrades are treated as change events requiring re-validation — not as a config swap.

6. Kill switches and rollback

Every autonomous action class has a documented blast radius, a circuit breaker triggered by anomaly thresholds, and a rehearsed rollback. If you cannot describe how to stop the agent in one sentence, it is not ready for production.

The commercial case is strongest where volume is high, rules are dense, and the action is reversible. That is administrative healthcare — an area where the CAQH Index estimates U.S. healthcare avoided roughly $258 billion in administrative cost through electronic transactions and improved data exchange, while still identifying substantial remaining automation headroom.[8]

Payers and TPAs. Prior authorization intake and documentation sufficiency checks, claims triage and pended-claim resolution, appeals and grievance packet assembly, provider data maintenance. Deloitte research reported by Becker's found the overwhelming majority of health plan executives expect AI to ease prior authorization burden[10] — but the AMA data above[7] means the only defensible design is one where the agent accelerates approvals and evidence gathering while adverse determinations remain with licensed reviewers, fully attributed.

Providers and RCM. Eligibility and benefit verification, coding support with citation to source documentation, denial root-cause classification and appeal drafting, referral routing, and inbound fax/document normalization — arguably the single highest-ROI agentic workload in the provider space.

Life sciences and labs. Protocol data reconciliation, result routing and exception handling, specimen and order matching across disparate LIS/EHR interfaces.

What is not ready: autonomous clinical decisions, unattended adverse determinations, and anything where the action cannot be reversed and the rationale cannot be reconstructed. The peer-reviewed adoption literature is consistent on this — health systems report enthusiasm alongside real challenges in validation, governance, and integration.[9] Believe both halves of that sentence.

Brandywine Consulting Partners is expanding its agentic AI practice deliberately, and in the order that makes deployments survive:

• ▸Transaction Layer Integration first. We act as the transactional connection point between client data and downstream systems, with zero-retention PHI handling and provenance on every write. Agents plug into a substrate that is already auditable. See Transaction Layer Integration.
• ▸Agent orchestration engineering. Supervisor/worker decomposition, typed tool contracts against FHIR, HL7v2, and X12, idempotent staged writes, and circuit breakers — implemented as infrastructure, not prompt engineering.
• ▸Governance and evaluation tooling. NIST AI RMF-aligned control mapping,[4][5] step-level eval harnesses, drift monitoring, and HTI-1-style transparency artifacts[6] produced as a deliverable rather than reconstructed during an audit.
• ▸Phased autonomy rollout. The L0–L4 ladder above, with promotion gated on measured reliability and signed off by a named governance owner.
• ▸Interoperability depth. Because the hard part of agentic AI in healthcare has never been the model. It is the integration surface — which is the work we have been doing all along.

Our position, plainly stated: agentic AI will deliver enormous value in healthcare operations over the next thirty-six months, and the organizations that capture it will be the ones that treated it as an integration and governance program with a model inside it — not an AI program with integration attached.

1. ▸Inventory reversible, high-volume workflows. Rank by volume × rule density × reversibility. Start where a mistake is recoverable.
2. ▸Stand up the transaction layer before the agent. Idempotency, provenance, staged writes, no PHI at rest.
3. ▸Define the autonomy ladder for your organization and assign a named owner to each promotion gate.
4. ▸Build the eval harness before the first pilot. If you cannot measure step-level reliability, you cannot promote anything.
5. ▸Map controls to NIST AI RMF and your HITRUST program now,[4][5] while the scope is one workflow instead of twelve.

We build the integration and governance substrate that makes agentic AI survivable — and then we build the agents on top of it. If you are evaluating an agentic initiative, or recovering one that stalled, we can help you assess where it sits on the autonomy ladder and what it needs to reach production.

• ▸Explore Transaction Layer Integration
• ▸Revisit the original analysis: The AI Agent Graveyard
• ▸Contact BCP to discuss your agentic AI roadmap

1. ▸Meyerson, E., Paolo, G., Dailey, R., et al. Solving a Million-Step LLM Task with Zero Errors. Cognizant AI Lab / UT Austin. arXiv:2511.09030. https://arxiv.org/html/2511.09030v1
2. ▸Fan, S., Ye, X., Huo, Y., et al. AgentProcessBench: Diagnosing Step-Level Process Quality in Tool-Using Agents. arXiv:2603.14465. https://arxiv.org/pdf/2603.14465v1
3. ▸Khanal, A., Tao, Y., Zhou, J. Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents. Northern Kentucky University. arXiv:2603.29231. https://doi.org/10.48550/arxiv.2603.29231
4. ▸National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1). https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.600-1.pdf
5. ▸National Institute of Standards and Technology. AI Risk Management Framework. https://www.nist.gov/itl/ai-risk-management-framework
6. ▸Office of the National Coordinator for Health IT. HTI-1 Final Rule — Decision Support Interventions and Predictive DSI Source Attributes. https://healthit.gov/regulations/hti-rules/hti-1-final-rule/
7. ▸American Medical Association. 2024 AMA Prior Authorization Physician Survey and accompanying release on AI and prior authorization denials. https://www.ama-assn.org/press-center/ama-press-releases/physicians-concerned-ai-increases-prior-authorization-denials
8. ▸CAQH. 2025 CAQH Index: U.S. Healthcare Avoided $258 Billion Through Automation, Interoperability and AI Adoption. https://www.globenewswire.com/news-release/2026/02/19/3241072/0/en/2025-CAQH-Index-Shows-U-S-Healthcare-Avoided-258-Billion-and-Accelerated-Automation-Interoperability-and-AI-Adoption.html
9. ▸Poon, E.G., Lemak, C.H., Rojas, J.C., Guptill, J., Classen, D. Adoption of artificial intelligence in healthcare: survey of health system priorities, successes, and challenges. J Am Med Inform Assoc. 2025;32(7):1093–1100. https://pmc.ncbi.nlm.nih.gov/articles/PMC12202002/
10. ▸Becker's Payer Issues, reporting Deloitte research: 93% of health plan executives expect AI to ease prior authorization. https://www.beckerspayer.com/research-analysis/93-of-health-plan-executives-expect-ai-to-ease-prior-authorization-deloitte/

Brandywine Consulting Partners provides integration, analytics, and AI engineering services to healthcare organizations. This article is technical and strategic guidance, not legal or compliance advice.