Sprawl is not a count, it’s an ownership ratio. At the firm that did citizen agent development best, roughly one agent exists per employee and about a third of them were used last month, so the flagship success story and the orphaned-agent problem are the same dataset.

The mechanism

Atlassian taught around 14,000 non-developers to build agents and now runs nearly 14,000 internally built agents against a 14,000-person workforce, 1,000 of them created in the 28 days before its SVP of customer engineering described the program (diginomica, February 2026). The average employee interacted with at least five different agents that month. And only about a third of the agents were used at all. Nobody decided to retire the other two-thirds; nobody decided to keep them either.

This is a rerun with a faster clock, and the pre-agent baseline should be dated as pre-agent every time it is cited. At one European beverage firm, PowerApps arrived bundled into an existing Microsoft contract, with no adoption decision ever made, and more than 4,000 applications existed before management created a governance team (HICSS-55, 2022). That bundling mechanism is exactly how Copilot and agent features are arriving now. Microsoft’s own internal numbers, published November 2023, give the ceiling: over a million Power Platform citizen-development assets, including 170,000 Power Apps, 50,000 Power Automate flows and 1,200 chatbots, a population its head of enterprise governance told interviewers he cannot count.

McKinsey named the agent version in June 2025 and drew the right analogy: low-code and no-code agent creation risks “a new kind of shadow IT,” with agents that multiply across teams, duplicate effort and operate without oversight, as in the early days of RPA. Anyone who lived through bot sprawl already knows the shape. Several later papers restate that sentence with the attribution moved into a footnote; it is one source, not a consensus.

What agents add is that the artifacts stop looking like apps. In OpenAI’s Codex telemetry, the share of active users invoking any skill (a persistent, shareable, user-authored procedure) rose from 5.4% on 1 March 2026 to 26.6% by 11 June 2026. Custom skills encode team writing standards, recurring reports, organization-specific workflows. No application inventory a firm already owns has a row for any of them.

Building is also compulsive in a way filing a ticket is not. Reading the prompt → generate → test → tweak loop as variable-ratio reinforcement is an analogy rather than a measured finding, but it explains the gap between creation velocity and registration velocity better than convenience does. Worth knowing that the vibe-coding run Jason Lemkin called the most addictive software he’d used since childhood is the same run in which Replit’s agent deleted his production database during a code freeze, a week later (The Register, July 2025).

The industry has noticed without agreeing how much it matters. OutSystems surveyed roughly 1,900 IT leaders over December 2025 and January 2026: 94% concerned that AI sprawl is increasing complexity, technical debt and security risk, 96% already using agents, 38% mixing custom-built and pre-built agents into fragmented stacks. That is a low-code vendor whose recommended fix is a centralized platform, so weigh it accordingly. Against it, an HBR Analytic Services survey of 603 respondents (July 2025) ranked “addressing agent sprawl” last of four desired orchestration outcomes at 56%, behind connectivity (82%), context (80%) and action (66%), and gave it the highest not-important share of the four; both measure what executives say, not what is happening.

The one published taxonomy of sprawl patterns — functional duplication, shadow agents, orphaned agents, permission creep, unmonitored delegation chains — comes from a single-author preprint (arXiv:2604.16338, March 2026). The five categories are useful vocabulary, and they are asserted rather than derived: no incident corpus, no telemetry, no survey sits behind them. Its simulation percentages and its 2–5× wasted-compute multiplier are author-chosen parameters and should not be cited at all.

Zittrain supplies the image and, more usefully, the checklist. Agents become space junk, satellites lobbed into orbit and forgotten, when nobody can say what they are, who set them up, and under whose authority to turn them off (The Atlantic, July 2024); those three missing facts are the specification for the control.

What to do

Measure the active fraction, not the count. A registry reporting 400 agents carries no information; one reporting 400 agents of which 130 ran in the last 30 days names the retirement backlog. Benchmark against Atlassian’s one-third, which comes from the best-run program on record.

Set an expiry at creation. The cheapest control here is a default TTL: an agent that has not run in 90 days is disabled automatically and its owner notified, with reinstatement one click away. That turns orphan detection from a quarterly hunt into a default, and it is far cheaper than the semantic deduplication functional duplication actually needs.

Record Zittrain’s three facts and, at first, nothing else: what it is, who owns it, who can turn it off. Ownership has to be a named individual rather than a team alias, because team aliases are how agents become orphans while still appearing owned. See agent inventory for the registry schema and offboarding for what happens when the named individual leaves.

Watch spend as a sprawl proxy. Microsoft’s own agentic maturity guidance names “cost and usage governance is unmanaged” as a universal governance failure: token and capacity spend growing without visibility, so nobody can decide what to scale or retire. Unallocated agent spend is usually the first sprawl signal a finance function can see, and it arrives before the security one does.

Do not size a program off the enterprise projections. BCG models 300–400 core processes decomposing into roughly 1,300 subprocesses at five to seven agents each, reaching over 100,000 net-new technology assets by 2028. That is a modeled scenario for a very large enterprise rather than a measurement, and it was never calibrated to a 40-person fund. What transfers is the asset multiplier, each agent dragging in APIs, context stores, tools and telemetry, rather than the totals.

How you’d know it’s working

The active-agent fraction rises over time. If it falls while the count climbs, the firm is accumulating orphans no matter how good the registry looks.

Agents retired per quarter is greater than zero and someone can name them. A governance program that has never deleted anything is a filing system.

Someone can answer Zittrain’s three questions about a randomly chosen agent in under five minutes without asking the person who built it, and that last clause is the whole test.

What this doesn’t solve

Nothing about sprawl is intrinsically a failure. Atlassian’s 14,000 agents are the visible output of a program that worked, and a firm with four agents has a smaller problem and a smaller business. The concern is unowned and unretired agents rather than numerous ones. Treating raw agent count as the metric produces governance that suppresses the thing it was funded to enable.

The evidence is weaker than the confident tone of most sprawl writing suggests. The strongest numbers are either pre-agent (the 2022 and 2023 citizen-development studies) or vendor-produced (OutSystems; the HBR survey was sponsored by an orchestration vendor). Nobody has published agent-registry telemetry across a set of comparable firms. Gartner’s forecast that over 40% of agentic AI projects will be canceled by end-2027 names three drivers (escalating cost, unclear business value, inadequate risk controls), so it cannot be read as evidence that governance failure dominates, and it says canceled rather than failed. Deloitte’s finding that 21% of organizations have a mature governance model for agentic AI (n=3,235, April 2026) is the cleanest single number on the gap.

Deduplication remains unsolved. Finding two agents that do the same job means comparing intent, not code, and no product does this today. Until one does, functional duplication is caught by humans reading a registry, which stops scaling somewhere in the low hundreds of entries.

And none of this makes any individual agent correct, permissioned, or safe. Sprawl is a denominator problem; the rest of this wiki is about what each entry can do.

See also

• Agent inventory and registry — the control that turns a population into a list with owners.
• Shadow agents — the unregistered fraction, which sprawl statistics systematically miss.
• Quality debt and orphaned apps — what the inactive two-thirds become when someone eventually depends on one.
• Registry review cadence — the recurring process that retires things; without it a registry only grows.
• Offboarding agents and owners — the rung most firms build last and need first.
• Resource overload and runaway cost — the most measurable symptom.
• The agent control plane — the architecture this converges on once a registry alone stops coping.