Leadership1 publisher3 min readPublished
Dynatrace's CTO argues instrumentation caps how fast agents can safely ship code
The CTO of Dynatrace argues that agents cannot be trusted with production work inside systems that do not report on themselves, and he cites two 2026 surveys of enterprise visibility to size the problem.
The Board Room · Leadership desk
What happened
- A January 2026 Neurones IT Asia report cited in the column finds that fewer than one in ten enterprise applications is fully observable today.
- A March 2026 SolarWinds survey put 77% of responding IT teams short of full visibility across their hybrid environments.
- Dynatrace's CTO and founder argues that as agents create, deploy and operate code, that visibility gap becomes a hard ceiling on how fast any organisation can safely scale.
- His own example chains ten agents at 95% accuracy each and reports cumulative accuracy dropping to roughly 60%.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- constraint Each agent added to a sequence raises the accuracy every other agent has to hit for the whole run to be trustworthy, so lengthening an agent pipeline is a reliability decision before it is a capability one.
- decision The choice a CIO makes this quarter is which applications emit enough telemetry to receive agent-written changes at all, and funding that telemetry competes with funding the agents.
- exposure Agents let loose in applications outside the observable set fail where no one is watching, and the first parties to read the failure are customers, regulators or auditors.
- contradiction The single maturity metric on offer is stated in two directions at once, so anyone adopting it has to decide first which way it counts before it can appear on a board slide.
Take the chain arithmetic first, because it is the part of the argument that does not depend on anyone's product. The column's example is ten agents at 95% accuracy each, run in sequence, with cumulative accuracy falling to roughly 60% [7]. Ten agents at 95% accuracy each does compound that way: 0.95 to the tenth power is 59.9% [8]. Run the requirement backwards and the bar gets specific. For a ten-step chain to finish correctly 95% of the time, each step has to be right about 99.5% of the time [9]. The multiplication assumes each error is independent and that nothing downstream catches it.
Bernd, the author, is the CTO and founder of Dynatrace, a unified observability and security company [3], and his conclusion is that observability becomes a precondition for moving past pilots [4]. The visibility figures he cites are not his company's: the one-in-ten number is attributed to a January 2026 Neurones IT Asia report [1], and the 77% figure to a March 2026 SolarWinds survey [2].
What the record does not carry is the link between those two figures and agents. Both measure the human-era estate, application observability and visibility across hybrid environments, and the column defines neither "fully observable" nor the sample or method behind either study [19]. So the ceiling is an inference. It is a reasonable one, and Bernd frames it as "speed and blindness at the same time" [6]: agents can write, deploy and optimise software faster than any human team, but they have no awareness of what happens once that software is running, and cannot see whether a change caused a slowdown, triggered an error cascade or violated a compliance boundary [5].
The recommended sequence is conventional. Start with innovation teams and lower-risk applications, prove the model, build institutional confidence, then let that trust expand to business-critical workloads [10]. The constraint on that plan is not appetite. More than nine in ten applications sit outside the fully observable set [17], and existing systems carry compliance obligations, integration dependencies and technical debt that cannot be automated away overnight [11]. The applications with usable telemetry get agents first. Those are not always the applications where the payback is largest.
One consequence gets a single line in the column and deserves more room. Intellectual property stops living in the codebase and starts living in the spec, because the code itself becomes a generated output [12]. If that holds, the artefact a reviewer signs and an auditor asks for is the spec.
The measurement advice is where the column is hardest to act on. It names the percentage of completions that require no human intervention as the metric gaining traction among teams running agent-led operations [13], and then adds: "The lower that number, the more effectively the AI is working" [14]. Those two sentences point in opposite directions.
What to watch
- Whether either cited study is published with sample size, method and a working definition of "fully observable".