Why Adding Agents Makes Deviation Investigations Worse

Deviation investigation is the highest-volume, highest-friction process in a GMP quality unit and the obvious place to point an AI agent. It is also, on the evidence published in the last nine months, close to the worst possible fit for a multi-agent architecture. The reason is structural rather than regulatory: an investigation is a chain in which each finding constrains the next question, and controlled evaluations show chained reasoning is exactly what breaks when you distribute work across agents.

The numbers are specific. In a controlled comparison of four coordination architectures building the same seven-service backend on the same model and the same $50 budget, a single agent produced the correct outcome in 28 of 28 attempts. A hierarchical organisation, one agent assigning work to others, managed 64%. A stigmergic swarm of eight concurrent agents managed 32%. An eleven-stage gated pipeline never produced a good outcome at all — it spent its entire project budget on five planning stages without writing a single line of implementation code. That work is Jeremy McEntire's The Organizational Physics of Multi-Agent AI, reported in Genetic Engineering & Biotechnology News on 11 March 2026 and CIO on 17 March 2026.

The manufacturing risk McEntire names is precise, and it is not hallucination. Multi-agent systems, he argues, "tend to optimize for internal process compliance rather than external outcome accuracy" — in a batch context, a coordination layer that approves a deviation, flags an issue as resolved or clears a batch record because the documentation satisfies the channel rather than because the issue is actually resolved. That is a defect no reviewer catches by reading the record, because the record looks right.

What the controlled evidence actually shows

The strongest dataset is not the software experiment. It is Towards a Science of Scaling Agent Systems, posted to arXiv on 9 December 2025 by a Google Research, MIT and University of Washington group and summarised on the Google Research blog on 28 January 2026. Its current version reports 260 configurations across six agentic benchmarks, five canonical architectures — single-agent plus independent, centralised, decentralised and hybrid multi-agent — and three model families, with tools, prompts and token budgets standardised so architecture is the variable rather than implementation.

Relative to the single-agent baseline, performance ranges from +80.8% on decomposable financial reasoning to −70.0% on sequential planning. The earlier 180-configuration version of the same study broke the planning degradation down by architecture: independent −70.0%, centralised −50.4%, decentralised −41.4%, hybrid −39.0%. Every multi-agent variant lost, and the spread is the 39–70% band now widely quoted. The same work reports error amplification of 17.2× for independent architectures against 4.4× where a centralised verifier exists.

Two things follow. The gains are real but task-shaped: parallelisable work benefits, chained work does not. And the mechanism is legible — a single agent runs a direct execution path, while coordination introduces branching, backtracking and consensus overhead that fragments a linear reasoning chain under a fixed compute budget. Each handoff also replaces the full record with a summary, and the detail that goes missing is the detail nobody knew was load-bearing.

McEntire's research note puts a number on the tax: token governance overhead scales with the links between agents rather than with capability, from 25.5% at three agents to 69.4% at fifteen. Two-thirds of the budget spent on coordinating.

Why a deviation investigation is the wrong shape for a swarm

Map an investigation onto the task taxonomy honestly. It opens with a discrepancy. The first question — isolated event or repeat — determines whether you pull a trend, and the trend result determines whether you go to the equipment, the material lot, the operator or the method. Root cause narrows by elimination. Scope of impact, the part FDA cites most, depends entirely on what root cause turned out to be, because it decides which other batches are implicated. CAPA depends on scope; effectiveness checks depend on CAPA.

That is a chain of dependent decisions with strong path dependency — the task class where every multi-agent architecture in the Google-led study lost between 39% and 70%. It is not a decomposable research problem where five agents each pull a source and a synthesiser merges the results. There are decomposable sub-tasks inside it — deviation history for the same equipment, the last three annual product reviews, environmental monitoring data for the affected room — and those are where a second agent earns its place. The chain itself is not one of them.

The failure mode inspectors already cite

FDA does not need a new AI-specific rule to catch this, because the defect it produces is one it cites constantly. 21 CFR 211.192 requires a thorough investigation of any unexplained discrepancy or failure of a batch to meet specifications, whether or not the batch has already been distributed, with a written record including conclusions and follow-up. The 2026 warning letters show what "thorough" means in practice, and it is never about speed.

CareFusion 213, LLC was cited under 211.192 on 30 April 2026 after more than 2,500 customer complaints since September 2023 covering foreign matter, missing components, empty vials and compromised seals. Medline Inc was cited on 28 May 2026 for failing to adequately investigate and implement CAPA after isolating objectionable microorganisms including Bacillus cereus from finished product on approximately nine occasions between June 2023 and August 2025. International Medication Systems Limited was cited on 2 July 2026 over multiple too-numerous-to-count microbial findings during manufacture of an epinephrine injection lot, including an ISO 5 viable air plate with confluent growth and insect larvae.

In each case records existed. The investigations were written, closed and filed. What FDA objected to was that they did not find the cause, did not extend to the batches the cause implicated and did not recur-proof the process. That is the defect a coordination layer produces, arrived at by a different route. An orchestrator that marks a stage complete when its acceptance criteria are met is doing what an under-resourced quality unit does when it closes a deviation to hit a thirty-day SOP commitment: satisfying the process rather than the question.

In Europe, EU GMP Chapter 1 clause 1.4(xiv) requires appropriate corrective and preventive actions to be identified and taken in response to investigations, with effectiveness monitored in line with quality risk management. Note the status: a revised Chapter 1 aligning to ICH Q9(R1) went out for stakeholder consultation on 3 September 2025, closing 3 December 2025. That consultation page reads "Closed" and the revised text has not been adopted, so the currently published Chapter 1 binds.

What is binding on 30 August 2026, and what is not

This is where most vendor material about agentic deviation management goes wrong, and the distinction is worth getting right before any of it reaches a validation plan. The wider enforcement record is in the companion piece on what regulators have actually approved, cited and refused in AI-enabled manufacturing.

EMA's Inspectors Working Group work plan targets Q4 2026 for delivering a final Annex 22 text to the European Commission. That is a target. No implementation period and no effective date have been announced, and Annex 22 does not appear in EudraLex Volume 4. Anyone writing "Annex 22 requires" in a URS in August 2026 is writing about a document that is not law.

The direction of travel is visible all the same. EMA ran a multistakeholder workshop on 30 June and 1 July 2026 — an open expert day, then a closed drafting-group day — explicitly to consider whether and under what safeguards dynamic, adaptive and probabilistic models, generative AI included, could be permitted in GMP applications. One of the six topics was framed as: what level and form of human-in-the-loop oversight is still required when guardrails are implemented. The prohibition may soften. The oversight question will not.

Separately, the EU AI Act deferrals are now law rather than proposal: the Digital Omnibus on AI moved Annex III standalone high-risk systems from 2 August 2026 to 2 December 2027 and Annex I product-embedded systems to 2 August 2028, per the Cloud Security Alliance's July 2026 analysis. Prohibited practices, the Article 4 AI literacy duty and GPAI obligations were already live before that, and Article 50 transparency applied from 2 August 2026 on the original schedule. Most GMP deviation AI is not Annex III high-risk in any case; the sectoral GMP stack is what governs it.

Where the 50–70% time saving comes from

Compilations of AI-assisted deviation case studies commonly report 50–70% reductions in investigation time. Treat that band as directional: it aggregates vendor-published cases rather than controlled trials and is not independently replicated. The mechanism behind it is real, and worth separating from the architecture question, because the saving mostly does not come from reasoning at all.

It comes from retrieval and drafting. McKinsey estimates roughly 30% of pharma staff time goes on documentation — dossiers, batch records, logs. A QA manager survey compiled by Climet put the average labour cost of a failure investigation at about $14,000 in 2023, up from roughly $10,000 in 2018. Typical SOP commitments run to thirty days for closure, strong sites clearing minor deviations inside fifteen — practice norms, not requirements, and no regulator sets either number.

Strip that apart and the honest claim is narrower than the headline. A model that pulls every prior deviation on the same equipment train, extracts the relevant environmental monitoring window and drafts a narrative once a human has settled the conclusion removes hours of retrieval and typing. It does not determine root cause, and the moment you build an architecture in which it appears to, you have moved a non-deterministic component into the position draft Annex 22 reserves for static deterministic models and put a GMP decision behind a summarisation boundary.

What this means in practice

Use the fewest agents that finish the task. McEntire's stated design rule, and the one the scaling data supports. One agent holding the entire investigation context beats three passing summaries. Add a second agent only for genuinely independent retrieval, and give it a verifier — centralised architectures contained error amplification to 4.4× against 17.2× for independent ones.

Anchor metrics to batch outcomes, not cycle time. A dashboard measuring days-to-close and percentage of deviations triaged by AI builds the exact incentive the coordination layer already has. Measure recurrence rate of the same root cause code, share of investigations where scope of impact was later widened, right-first-time on batch records, and how often a human reviewer overturned the AI's proposed cause. That last one is your model performance metric and your validation evidence at once.

Draw the critical line explicitly in the risk assessment. Retrieval, trend surfacing and narrative drafting on a settled conclusion sit outside critical GMP applications. Root cause, scope of impact and batch disposition sit inside. Write that boundary into the intended-use statement and the URS, validate against the 2011 Annex 11 that binds today, and design to survive the draft that does not.

Make the human step real. A reviewer who sees only the agent's summary is not oversight, they are a signature. Give the qualified reviewer the records the model used, log which were retrieved and which were not, and make the audit trail resolve every model-influenced write to an identified system account plus an accountable human. The current Annex 11 already expects that; the draft revision would add creation events and immutability, as set out alongside the enforcement and approval record for AI in manufacturing.

Budget for the coordination tax. At three agents roughly a quarter of tokens go on governance; at fifteen, nearly 70%. That is a cost line and a latency line, and it belongs in the business case next to the $14,000 labour figure rather than being discovered after go-live.

The pitch for agentic deviation management is that more agents mean more thoroughness. The published evidence says the opposite for the part of the work that matters, and the failure it produces reads as a clean record and lands as a warning letter.