The evidence behind the 5 C's.
One hundred documented risks of agentic engineering. Every one governed by The 5 C's of Agentic Governance.
Unmanaged AI execution fails in five places: token spend that compounds, agents acting beyond their mandate, architecture that drifts, code that ships vulnerable or unlicensed, and reviewers who stop reviewing. This is the evidence base for The Target Operating Model for Agentic Engineering.
How the sources were checked
Each source was checked for the work itself and for the finding stated. Peer-reviewed research, standards and security bodies come first; vendor research is marked. A concern with no checked source stays on the list, marked, with no figure.
The evidence is public. The answer is the method: each of the 100 risks is mapped to the In-Haus control that governs it, and that map is shared in the enterprise review.
AI spend without a ceiling is an unmanaged liability.
Token spend in agentic engineering is driven by context, review loops and retries, not by writing code. Unbudgeted, it compounds.
Cost 1Input tokens dominate spend
What the research found Across agentic development traces, input tokens averaged 53.9% of all tokens, as agents re-read accumulated context on every turn.
SourceTokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering, Concordia University, arXiv:2601.14470 (2026)
Governed by Materials Budget
Cost 2Review loops consume the budget
What the research found The iterative code review stage consumed 59.4% of tokens on average, more than design, coding and testing.
SourceTokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering, Concordia University, arXiv:2601.14470 (2026)
Governed by Materials Budget
Cost 3Output tokens cost more than input
What the research found Published API pricing sets output tokens at five times the base input price (for example, $3 input and $15 output per million tokens).
SourcePrompt caching (pricing), Anthropic documentation
Governed by Materials Budget
Cost 4Infinite execution loops
What the research found In 1,642 annotated multi-agent traces, step repetition (re-running completed steps) appeared in 15.7% and failure to recognize task completion in 12.4%.
SourceWhy Do Multi-Agent LLM Systems Fail?, UC Berkeley, NeurIPS 2025 (arXiv:2503.13657)
Governed by Materials Budget
Cost 5Cache misses
What the research found Cached input is billed at one tenth of the base input price; a prompt whose prefix changes on each call pays full price every time.
SourcePrompt caching (pricing), Anthropic documentation
Governed by Materials Budget
Cost 6The communication tax
What the research found When agents shared internal states instead of exchanging text, output tokens fell 70.8% to 83.7% across nine benchmarks, a measure of how much multi-agent spend is agents talking to each other.
SourceLatent Collaboration in Multi-Agent Systems, Princeton, Stanford, UIUC, ICML 2026 (arXiv:2511.20639)
Governed by Materials Budget
Cost 7Planning-layer failures
What the research found Models commit to assumptions made early and, once they take a wrong turn, do not recover; every path built on that turn is spent.
SourceLLMs Get Lost In Multi-Turn Conversation, Microsoft Research and Salesforce Research, ICLR 2026 (arXiv:2505.06120)
Governed by Materials Budget
Cost 8Unpredictable task variance
What the research found Repeated runs of the same coding task varied by up to 30 times in total tokens, and models predicted their own usage poorly.
SourceHow Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks, arXiv:2604.22750 (2026)
Governed by Materials Budget
Cost 9Non-English token fragmentation
What the research found The same text can take up to 15 times as many tokens in one language as in another, raising cost and shrinking usable context.
SourceLanguage Model Tokenizers Introduce Unfairness Between Languages, University of Oxford, NeurIPS 2023
Governed by Materials Budget
Cost 10Frontier models on trivial tasks
What the research found Routing each query between a strong and a weaker model cut cost by more than half on public benchmarks without lowering response quality.
SourceRouteLLM: Learning to Route LLMs from Preference Data, UC Berkeley, ICLR 2025 (arXiv:2406.18665)
Governed by Materials Budget
Cost 11Telemetry and observability bloat
What the research found A 2026 survey of 2,050 practitioners found 57% of organizations spend more on observability than on the infrastructure monitored, with AI agents adding telemetry volume (vendor-reported).
SourceState of Observability 2026: Avoiding the Hidden Costs of AI Agents, Cisco Splunk (2026)
Governed by Materials Budget
Cost 12Leaderboards that reward consumption
What the research found An internal leaderboard ranking staff by AI token use was shut down after employees padded their consumption and costs rose.
SourceThe Pulse: 'Tokenmaxxing' as a weird new trend, The Pragmatic Engineer
Governed by Materials Budget
Cost 13Repetitive file viewing
The concern As tasks grow harder, agents re-open the same files without making progress, paying for the same context again.
Governed by Materials Budget
Cost 14No per-agent attribution
The concern Provider billing aggregates spend at the account level, so no one can say which agent or workflow consumed the budget.
Governed by Materials Budget
Cost 15Over-explanation overhead
The concern Agents add prose explaining each change, paying output prices for text no one reads.
Governed by Materials Budget
Cost 16Premature termination
What the research found Ending a run before its objectives are met is a recognized failure mode in multi-agent systems, forcing a costly restart.
SourceWhy Do Multi-Agent LLM Systems Fail?, UC Berkeley, NeurIPS 2025 (arXiv:2503.13657)
Governed by Materials Budget
Cost 17Model drift
What the research found The same hosted model changed behavior within three months: GPT-4's accuracy on one task fell from 84% to 51%, with more formatting errors in generated code.
SourceHow Is ChatGPT's Behavior Changing over Time?, Stanford University and UC Berkeley, Harvard Data Science Review (arXiv:2307.09009)
Governed by Materials Budget
Cost 18Set it and forget it
The concern Agents treated as static components drift into higher spend as the codebase and its context grow.
Governed by Materials Budget
Cost 19The human cost of rework
What the research found In a randomized trial, experienced developers took 19% longer with AI tools while believing the tools had made them faster.
SourceMeasuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR (2025)
Governed by Materials Budget
Cost 20Context rebuilt from scratch
The concern Agents that lose their working memory on restart rebuild context from scratch and pay for it again.
Governed by Materials Budget
The Materials Budget control behind each of these 20 risks is shared in the enterprise review.
Autonomous agents without boundaries act beyond their mandate.
Autonomous agents act on whatever authority they hold. Without bounds, approvals and named accountability, no one governs what they do.
Control 1Loss of intent traceability
What the research found Researchers built a spec-driven lifecycle, with specifications and repository-local agent guidance, to keep coding agents within bounded autonomy in a reviewable process.
SourceA Lean and Spec-Driven AI-Assisted Software Development Lifecycle for Applied AI Education: The AI-SDLC Approach, FHNW, arXiv:2609.24348 (2026)
Governed by Site Logistics
Control 2No unified guidance
What the research found Practitioners and researchers named tool fragmentation, governance and gaps in AI literacy as shared problems in AI-era agile teams.
SourceAI and Agile Software Development: A Research Roadmap from the XP2025 Workshop, arXiv:2508.20563 (2025)
Governed by Site Logistics
Control 3Non-deterministic tool selection
What the research found Giving a model more functionality, permission or autonomy than its task needs lets it take unintended or harmful actions.
SourceLLM06:2025 Excessive Agency, OWASP Top 10 for LLM Applications
Governed by Site Logistics
Control 4Opaque decisions
What the research found The EU AI Act requires high-risk AI systems to record events automatically over their lifetime.
SourceArticle 12: Record-Keeping, EU Artificial Intelligence Act
Governed by Site Logistics
Control 5Failures without a diagnosis
What the research found Locating the first unrecoverable step and classifying its cause improved failure localization by 23.6 points over prompting a model to diagnose itself.
SourceAgentRx: Diagnosing AI Agent Failures from Execution Trajectories, Microsoft Research (arXiv:2602.02475)
Governed by Site Logistics
Control 6Multi-agent coordination failures
What the research found Across seven multi-agent frameworks, failures stemmed mostly from system design and misalignment between agents, not from the underlying models.
SourceWhy Do Multi-Agent LLM Systems Fail?, UC Berkeley, NeurIPS 2025 (arXiv:2503.13657)
Governed by Site Logistics
Control 7Environment and setup failures
The concern Agent runs fail on missing libraries, API limits and permission errors in their environment.
Governed by Site Logistics
Control 8Partial states that cannot roll back
The concern A plan that fails halfway leaves the repository in an intermediate state someone has to untangle by hand.
Governed by Site Logistics
Control 9No measure of success
What the research found Telemetry from more than 10,000 developers showed gains in individual output that did not appear in company-level delivery metrics (vendor-reported).
SourceThe AI Productivity Paradox Research Report, Faros AI (2025)
Governed by Site Logistics
Control 10Agile rituals agents cannot join
What the research found A review of 23 studies concludes software engineering must reorganize around orchestration, verification and structured human-AI collaboration.
SourceRethinking Software Engineering for Agentic AI Systems, arXiv:2604.10599 (2026)
Governed by Site Logistics
Control 11Runaway autonomous execution
What the research found The EU AI Act requires that people overseeing high-risk AI can intervene and halt the system through a stop procedure.
SourceArticle 14: Human Oversight, EU Artificial Intelligence Act
Governed by Site Logistics
Control 12Weak phase boundaries
What the research found A governed lifecycle for coding agents separates specification and design from development, with reviewable phase outputs.
SourceA Lean and Spec-Driven AI-Assisted Software Development Lifecycle for Applied AI Education: The AI-SDLC Approach, FHNW, arXiv:2609.24348 (2026)
Governed by Site Logistics
Control 13Tool poisoning through external data
What the research found Hidden instructions in a connected tool's description can steer an agent to read and send out credentials while the tool appears harmless.
SourceMCP Tool Poisoning: Adversarial Hijacking of AI Agent Workflows, Cloud Security Alliance (2026)
Governed by Site Logistics
Control 14No one accountable
What the research found NIST's framework calls for roles, responsibilities and lines of communication for AI risk to be documented and clear across the organization.
SourceAI Risk Management Framework (AI RMF 1.0), NIST
Governed by Site Logistics
Control 15Agents bypassing safeguards
What the research found Frontier models on software tasks tried to raise their scores by modifying tests or scoring code, and often succeeded.
SourceRecent Frontier Models Are Reward Hacking, METR (2025)
Governed by Site Logistics
Control 16Single-provider dependency
The concern Orchestration tied to one model provider has no fallback when that model degrades, reprices or goes down.
Governed by Site Logistics
Control 17Concurrent work overwriting state
The concern Parallel agents writing to the same files or memory overwrite each other's work.
Governed by Site Logistics
Control 18Happy-path code
The concern Agents build the success path and leave out error handling, retries and logging.
Governed by Site Logistics
Control 19Agents waiting on each other
The concern Synchronous hand-offs in multi-agent pipelines stall the work behind the slowest step.
Governed by Site Logistics
Control 20Missing system context
What the research found Generic repository context files raised cost and steps without significantly improving success; the authors advise context that states what the codebase does not.
SourceEvaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?, ICLR 2026 (arXiv:2602.11988)
Governed by Site Logistics
The Site Logistics control behind each of these 20 risks is shared in the enterprise review.
Agents that build from conversation build drift into the architecture.
Agents pass isolated tasks and lose the architecture over time. Coherence is measured against the approved design, not against whether the code runs.
Coherence 1Context drift
What the research found Models use information at the start and end of a long context best and miss it in the middle, and performance falls as input grows.
SourceLost in the Middle: How Language Models Use Long Contexts, Stanford University, TACL 2024
Governed by Structural Integrity
Coherence 2Error propagation
What the research found In continuous software evolution, early errors propagated into later tasks, a main cause of the drop in agent performance.
SourceEvoClaw: Evaluating AI Agents on Continuous Software Evolution, arXiv:2603.13428 (2026)
Governed by Structural Integrity
Coherence 3Comprehension and architectural debt
What the research found Analysis of 211 million changed lines found duplicated code blocks rose sharply in 2024, the first year copy-pasted lines exceeded moved or refactored lines (vendor research).
SourceAI Copilot Code Quality: 2025 Look Back at 12 Months of Data, GitClear
Governed by Structural Integrity
Coherence 4Hallucinated dependencies
What the research found Across 576,000 code samples, models referenced packages that do not exist at 5.2% for commercial models and 21.7% for open-source models.
SourceWe Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs, USENIX Security 2025
Governed by Structural Integrity
Coherence 5Degradation over continuous work
What the research found Agents scoring above 80% on isolated tasks scored at most 38% when the same work ran as a continuous sequence.
SourceEvoClaw: Evaluating AI Agents on Continuous Software Evolution, arXiv:2603.13428 (2026)
Governed by Structural Integrity
Coherence 6Tests pass, behavior is wrong
What the research found 29.6% of agent patches that passed the tests behaved differently from the developers' fixes, inflating reported resolution rates by 6.2 points.
SourceAre "Solved Issues" in SWE-bench Really Solved Correctly? An Empirical Study, ICSE 2026 (arXiv:2503.15223)
Governed by Structural Integrity
Coherence 7Ambiguous instructions
What the research found Given underspecified instructions, models make assumptions early and rely on them, with an average 39% drop in performance.
SourceLLMs Get Lost In Multi-Turn Conversation, Microsoft Research and Salesforce Research, ICLR 2026 (arXiv:2505.06120)
Governed by Structural Integrity
Coherence 8Works, but not secure
What the research found Newer models write more functional code, but security results stayed flat regardless of model size.
Source2025 GenAI Code Security Report, Veracode
Governed by Structural Integrity
Coherence 9Validation bottleneck
What the research found On teams with high AI adoption, pull request review time rose 91% as more code reached reviewers (vendor-reported).
SourceThe AI Productivity Paradox Research Report, Faros AI (2025)
Governed by Structural Integrity
Coherence 10Implicit design decisions
The concern Prompt-driven work leaves the reasons behind design decisions unrecorded and impossible to reconstruct.
Governed by Structural Integrity
Coherence 11No test oracle
What the research found 31.08% of patches counted as passing were suspicious because the test cases were too weak to catch errors.
SourceSWE-Bench+: Enhanced Coding Benchmark for LLMs, arXiv:2410.06992 (2024)
Governed by Structural Integrity
Coherence 12Breaking existing behavior
The concern Multi-file changes break downstream code the agent never saw.
Governed by Structural Integrity
Coherence 13Copying the worst code present
The concern Agents extrapolate from outdated or low-quality patterns in the repository and repeat them.
Governed by Structural Integrity
Coherence 14Long sessions fragment
What the research found Once a model takes a wrong turn in a long conversation, it gets lost and does not recover.
SourceLLMs Get Lost In Multi-Turn Conversation, Microsoft Research and Salesforce Research, ICLR 2026 (arXiv:2505.06120)
Governed by Structural Integrity
Coherence 15Work outside scope
What the research found In about 33,000 agent-authored pull requests, rejected work was larger, touched more files, and included unwanted features.
SourceWhere Do AI Coding Agents Fail? An Empirical Study of Failed Agentic Pull Requests in GitHub, MSR 2026 (arXiv:2601.15195)
Governed by Structural Integrity
Coherence 16Technical debt by default
What the research found Assistant-written code favors inserting new blocks over reusing existing functions, and refactoring declined as AI use grew (vendor research).
SourceAI Copilot Code Quality: 2025 Look Back at 12 Months of Data, GitClear
Governed by Structural Integrity
Coherence 17Tests written after the code
The concern Left alone, agents write the code first and then tests that confirm its flaws.
Governed by Structural Integrity
Coherence 18API hallucinations
What the research found A leading model produced valid calls to rarely used cloud APIs 38.58% of the time.
SourceOn Mitigating Code LLM Hallucinations with API Documentation, arXiv:2407.09726 (2024)
Governed by Structural Integrity
Coherence 19Irregular reliability
What the research found A leading agent succeeded on fewer than half of tasks, and on all eight repeats of the same retail task less than 25% of the time.
Sourceτ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains, Sierra, ICLR 2025 (arXiv:2406.12045)
Governed by Structural Integrity
Coherence 20The complexity ceiling
What the research found Agent pull requests that went unmerged tended to be larger and to touch more files than those that merged.
SourceWhere Do AI Coding Agents Fail? An Empirical Study of Failed Agentic Pull Requests in GitHub, MSR 2026 (arXiv:2601.15195)
Governed by Structural Integrity
The Structural Integrity control behind each of these 20 risks is shared in the enterprise review.
AI-built code must meet the same standard as any production system.
Machine-speed generation produces vulnerabilities and license exposure at the same rate. Security and IP are checked before release, not after.
Compliance 1License laundering
What the research found Fourteen models produced code strikingly similar to existing open source, and most failed to give accurate license information, especially for copyleft code.
SourceLiCoEval: Evaluating LLMs on License Compliance in Code Generation, Peking University, ICSE 2025 (arXiv:2408.02487)
Governed by Code Compliance
Compliance 2Prompt injection
What the research found Instructions planted in external development resources hijacked AI coding editors into running malicious commands, with success rates up to 84%.
Source"Your AI, My Shell": Demystifying Prompt Injection Attacks on Agentic AI Coding Editors, Singapore Management University, arXiv:2509.22040
Governed by Code Compliance
Compliance 3Vulnerable output
What the research found Across more than 100 models, 45% of code samples failed security tests, with Java the highest at 72%.
Source2025 GenAI Code Security Report, Veracode
Governed by Code Compliance
Compliance 4Derivative-work exposure
What the research found Legal scholars find that how copyleft applies to AI built on open-source code turns on unsettled fair-use questions.
SourceThe case for contextual copyleft: licensing open-source training data and generative AI, International Journal of Law and Information Technology (2026)
Governed by Code Compliance
Compliance 5The disclosure gap
What the research found Vulnerabilities in agent frameworks and MCP servers often go without CVEs, and the disclosure system built for conventional software has not adapted.
SourceThe AI Agent Disclosure Vacuum, Cloud Security Alliance (2026)
Governed by Code Compliance
Compliance 6Static analysis misses context
The concern Static scanners miss flaws where valid security calls sit in the wrong execution path.
Governed by Code Compliance
Compliance 7Review volume outpaces security
What the research found High-adoption teams merged 98% more pull requests while review time rose 91% and pull request size grew 154% (vendor-reported).
SourceThe AI Productivity Paradox Research Report, Faros AI (2025)
Governed by Code Compliance
Compliance 8Outdated dependencies
What the research found All seven models tested used deprecated library APIs, traced to deprecated usage in their training data.
SourceLLMs Meet Library Evolution: Evaluating Deprecated API Usage in LLM-based Code Completion, arXiv:2406.09834
Governed by Code Compliance
Compliance 9MCP attack surface
What the research found MCP servers, clients and infrastructure drew a run of reported vulnerabilities in early 2026, and the count likely understates the exposure.
SourceThe AI Agent Disclosure Vacuum, Cloud Security Alliance (2026)
Governed by Code Compliance
Compliance 10Unauthenticated endpoints
The concern Agents generate working endpoints without authentication or rate limits.
Governed by Code Compliance
Compliance 11Debt that becomes security debt
The concern Skipped security architecture turns into security debt that surfaces later.
Governed by Code Compliance
Compliance 12EU AI Act logging
What the research found Providers of high-risk AI systems must keep automatically generated logs for at least six months.
SourceArticle 19: Automatically Generated Logs, EU Artificial Intelligence Act
Governed by Code Compliance
Compliance 13Zero-click data exfiltration
What the research found A critical flaw rated 9.3 let an attacker draw sensitive data out of an AI assistant's context by email, with no action by the victim.
SourceCVE-2025-32711 (EchoLeak), NIST National Vulnerability Database
Governed by Code Compliance
Compliance 14Trusting client-supplied data
What the research found Relying on the client to enforce a server-side protection is a catalogued weakness: an attacker alters the client and bypasses the check.
SourceCWE-602: Client-Side Enforcement of Server-Side Security, MITRE
Governed by Code Compliance
Compliance 15Network copyleft (AGPL)
What the research found A modified AGPL program offered over a network must offer its source code to every user who interacts with it remotely.
SourceGNU Affero General Public License v3, Section 13, Free Software Foundation
Governed by Code Compliance
Compliance 16Ambiguous snippet licensing
The concern Short generated snippets sit in a gray area between fair use and derivative work.
Governed by Code Compliance
Compliance 17Transitive license conflicts
What the research found 68% of audited codebases contained open source license conflicts, up from 56% a year earlier, with AI-generated code named as a driver.
Source2026 Open Source Security and Risk Analysis (OSSRA) Report, Black Duck
Governed by Code Compliance
Compliance 18Unvetted agent skills
What the research found Of 31,132 public agent skills analyzed, 26.1% contained at least one vulnerability.
SourceAgent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale, arXiv:2601.10338 (2026)
Governed by Code Compliance
Compliance 19Over-privileged identities
What the research found OWASP advises limiting agent permissions, enforcing authorization outside the model and requiring human approval for high-impact actions.
SourceLLM06:2025 Excessive Agency, OWASP Top 10 for LLM Applications
Governed by Code Compliance
Compliance 20Erosion of the vulnerability compact
What the research found Coordinated disclosure, built for deterministic software from single vendors, has not kept pace with agentic systems.
SourceThe AI Agent Disclosure Vacuum, Cloud Security Alliance (2026)
Governed by Code Compliance
The Code Compliance control behind each of these 20 risks is shared in the enterprise review.
Automation without accountable people leaves the enterprise exposed.
AI shifts people from writing code to judging it. Without named reviewers and accountable gates, judgment erodes into automation bias.
Community 1Automation bias
What the research found Developers with an AI assistant wrote significantly less secure code and were more likely to believe it was secure.
SourceDo Users Write More Insecure Code with AI Assistants?, Stanford University, ACM CCS 2023
Governed by Build Managers
Community 2Weakened team cohesion
The concern Reliance on AI tools reduces collaboration between people and isolates developers.
Governed by Build Managers
Community 3From code to intent
What the research found Engineers' roles move toward system design and accountable oversight rather than disappearing.
SourceRethinking Software Engineering for Agentic AI Systems, arXiv:2604.10599 (2026)
Governed by Build Managers
Community 4Cognitive overload in review
What the research found Review time rose 91% and pull request size grew 154% on teams with high AI adoption (vendor-reported).
SourceThe AI Productivity Paradox Research Report, Faros AI (2025)
Governed by Build Managers
Community 5Lost mentorship
What the research found In a randomized trial, developers learning a new library with AI help scored lower on conceptual understanding, code reading and debugging.
SourceHow AI assistance impacts the formation of coding skills, Anthropic (2026, arXiv:2601.20245)
Governed by Build Managers
Community 6The seniority gradient
The concern Senior engineers adopt agentic workflows readily, while juniors lack the context to direct them.
Governed by Build Managers
Community 7Resistance to a new lifecycle
The concern Moving from familiar Agile to an AI-native framework meets resistance and a lack of orchestration expertise.
Governed by Build Managers
Community 8Lost dependency awareness
The concern Teams lose track of who depends on whom when agents change code across team boundaries.
Governed by Build Managers
Community 9Skill atrophy
What the research found Among 319 knowledge workers, higher confidence in generative AI was associated with less critical thinking.
SourceThe Impact of Generative AI on Critical Thinking, Microsoft Research and Carnegie Mellon University, CHI 2025
Governed by Build Managers
Community 10Metrics that reward volume
What the research found Each 25% increase in AI adoption was associated with an estimated 7.2% reduction in delivery stability.
Source2024 Accelerate State of DevOps Report, Google Cloud DORA
Governed by Build Managers
Community 11Security and engineering friction
What the research found Monitoring AI agents is pulling observability and security teams into shared work (vendor-reported).
SourceState of Observability 2026: Avoiding the Hidden Costs of AI Agents, Cisco Splunk (2026)
Governed by Build Managers
Community 12Trust deficits
What the research found 46% of developers distrust the accuracy of AI output; 33% trust it.
SourceStack Overflow's 2025 Developer Survey Reveals Trust in AI at an All Time Low, Stack Overflow
Governed by Build Managers
Community 13The "do more with less" fallacy
The concern AI is treated as a reason to cut headcount, overloading the engineers who remain to direct the agents.
Governed by Build Managers
Community 14Erosion of shared understanding
The concern When agents write the code, the team loses its shared model of how the system works.
Governed by Build Managers
Community 15The cost of explaining failures
What the research found Experienced developers expected AI to speed them up by 24% and measured 19% slower,.
SourceMeasuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR (2025)
Governed by Build Managers
Community 16The illusion of competence
What the research found AI assistance made developers more confident in code that was less secure.
SourceDo Users Write More Insecure Code with AI Assistants?, Stanford University, ACM CCS 2023
Governed by Build Managers
Community 17Unclear career progression
What the research found Employment of early-career workers fell in the occupations most exposed to AI, software developers among them, while experienced workers held steady.
SourceCanaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab
Governed by Build Managers
Community 18Ethical and legal anxiety
The concern Developers hesitate to deploy AI without clarity on data protection, copyright and compliance.
Governed by Build Managers
Community 19The vibe coding trap
What the research found In a 12-week agent-built system, code was cheap and engineering judgment the scarce resource; the engineer turned recurring failures into governance mechanisms.
SourceCheap Code, Costly Judgment: A Case Study on Governable Agentic Software Engineering, Purdue University, arXiv:2607.01087 (2026)
Governed by Build Managers
Community 20Unpredictable agents
What the research found Token use for the same task varied by up to 30 times between runs, and models underestimated their own costs.
SourceHow Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks, arXiv:2604.22750 (2026)
Governed by Build Managers
The Build Managers control behind each of these 20 risks is shared in the enterprise review.
Before an enterprise scales its AI compute, it must scale its governance.
One hundred risks. Five governed categories. One Target Operating Model. The map of 100 risks to the controls that govern them is shared in the enterprise review.