# AI Agents at Global Scale Meet Tribal Dungeons _Maersk's Dmitry Buykin says production agents fail in tribal dungeons, and fixing them took 100,000 corrections and a 20:1 SOP corpus._ **Published:** 2026-08-29 **Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/ai-agents-at-global-scale-meet-tribal-dungeons --- [AI agents at global scale](https://www.youtube.com/watch?v=dQ-_i1tZiws) do not fail on happy paths. They fail in what Maersk practitioner Dmitry Buykin calls tribal dungeons. Maersk agent deploymentCoreDmitry Buykin reports on shipping operations agent built for global scaleFrom the articleThe Maersk case suggests the moat in production agents is not the model but the SOP corpus and the correction loop that enterprises actually control.exposesTribal dungeons emergeDriverEdge cases where one step fails and legacy systems go incoherentFrom the article 2 mentionsThey fail in what Maersk practitioner Dmitry Buykin calls tribal dungeons.drivesLong tail dominates costsDriverException work by experts across incomplete legacy systems exceeds base workflowFrom the articleWhat remains is the long tail where happy paths break.demandsLoop harness encloses agentCoreSystem wraps agent with preconditions, back-end calls, validation, and evidenceFrom the article 2 mentionsSpikes and latencies range from a few minutes to 10 minutes, limited by legacy back ends, not the agent loop.enablesCorrections beat bigger modelsOutcome100,000 human corrections outperformed scaling up model parameters for edge casesFrom the articleThe Maersk case suggests the moat in production agents is not the model but the SOP corpus and the correction loop that enterprises actually control.Maersk agent deploymentCoreDmitry Buykin reports on shipping operations agent built for global scaleFrom the articleThe Maersk case suggests the moat in production agents is not the model but the SOP corpus and the correction loop that enterprises actually control.Screenshots fail as SOPsContextLegacy SOPs are click sequences lacking preconditions, validation, and recovery stepsFrom the articleBuykin said those legacy SOPs are a bunch of screenshots in sequence.exposesTribal dungeons emergeDriverEdge cases where one step fails and legacy systems go incoherentFrom the article 2 mentionsThey fail in what Maersk practitioner Dmitry Buykin calls tribal dungeons.drivesLong tail dominates costsDriverException work by experts across incomplete legacy systems exceeds base workflowFrom the articleWhat remains is the long tail where happy paths break.demandsLoop harness encloses agentCoreSystem wraps agent with preconditions, back-end calls, validation, and evidenceFrom the article 2 mentionsSpikes and latencies range from a few minutes to 10 minutes, limited by legacy back ends, not the agent loop.Corrections beat bigger modelsOutcome100,000 human corrections outperformed scaling up model parameters for edge casesFrom the articleThe Maersk case suggests the moat in production agents is not the model but the SOP corpus and the correction loop that enterprises actually control.Harness not freedomContextAgent constrained inside validated loop rather than given open-ended autonomyrequires20 to 1 corpus ratioContextEach corrected workflow requires twenty supporting SOP documents in the knowledge baseFrom the article 2 mentionsMaersk built three parts: SOP memory organized as a corpus, execution runtime, and feedback capture. Buykin detailed the work in a practitioner report for [AI Engineer](https://www.youtube.com/watch?v=dQ-_i1tZiws) on shipping operations at [Tribal Dungeons of Global Shipping: AI Agents at Global Scale](https://www.youtube.com/watch?v=dQ-_i1tZiws). Every shipment is an orchestration of many parallel state machines. ## The expensive part is the long tail The easy majority is already automated in many companies. What remains is the long tail where happy paths break. One step fails to complete, systems go incoherent, and an expert must orchestrate across multiple incomplete legacy systems. That exception work costs more than the base workflow. ## Screenshots are not a process Standard operating procedures in regulated industries capture what a person sees and clicks. Buykin said those legacy SOPs are a bunch of screenshots in sequence. An agent SOP needs preconditions, decisions, identifiers, back-end calls, validation, recovery and evidence of successful execution. Experts own the what. Agents own the how. The signal process depends on many systems being coherent at once. If any step cannot complete, an exception becomes a guardrail. ## The system is the loop around the agent Maersk built three parts: SOP memory organized as a corpus, execution runtime, and feedback capture. The agent loop is not the system. The refining loop around it is. The corpus is the asset. It is 20 to 1 versus runtime, modified for country conditions, and illustrated by an SAP slide where the same process means different things in different countries. Production runs over 200 instances concurrently. Spikes and latencies range from a few minutes to 10 minutes, limited by legacy back ends, not the agent loop. ## Why corrections beat bigger models Expert time is the bottleneck, so Maersk built a triage bench that clusters failures. The trace is shared evidence for experts and engineers to agree on what happened. A correction only counts when it becomes an executable change. That is the line between opinion and production fix. Quality comes from replaying real examples with disabled write rights and checking if behavior improved. Maersk logged over 100,000 corrections in the last 9 months. Heat maps turned thousands of traces into priorities. Buykin said turning one red block takes one to two months of effort from engineers and agents together. ## Harness for a cage, not freedom Discovery needs agent freedom. Production needs a cage. The harness makes dumb mistakes impossible. Wrong workflow triggers classifier eval. Wrong gate triggers right gate. Wrong assumption triggers review. Buykin said the team does not use MCPs because bloated system responses need distilling. They tune tools through function calling to control quality. His five moves: make work representable, make execution bounded, make behavior observable, make correction cheap, and make improvement compound. Successful sequences are merged into larger composite tools that roll out across hundreds of countries at once. ## Why this matters This flips the usual agent demo narrative. Model capability is not the constraint. Representing messy operational knowledge safely is. For enterprises, accuracy was earned one small correction at a time, not designed in one diagram. Pipe coding and spec driven development plateaued early, Buykin showed, and real gains came after. StartupHub.ai data shows You raised $80M in its 2023 Series A, and the competitors we track in that search and answer space include [Perplexity AI](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/perplexity-ceo-on-ai-boom-and-hybrid-local-future), [Google (NASDAQ:GOOGL)](https://www.google.com/finance/quote/GOOGL:NASDAQ), Lucidworks, matey and Dante. The Maersk case suggests the moat in production agents is not the model but the SOP corpus and the correction loop that enterprises actually control. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.