Fieldstate research · visual edition
2026
AI assistance is mainstream. Agent use is becoming common. High-authority agentic delivery is not. Thirty-nine pages on what that distinction costs you, and how to tell where your delivery system actually stands.
The central thesis
It is whether your delivery system can convert faster generation into safe, maintainable, valuable change, without overwhelming verification, weakening human capability, or increasing operational risk.
A system with broad permissions and weak controls is not mature. It is unsupervised.
A junior who can answer the one question is becoming senior.
A CTO who makes room for them to learn it is a leader worth following.
A practical test
Not a maturity scale. A test you can run today, on the next change in front of you.
Can we understand it?
Not "does it run" — can a human hold its shape in their head?
Can we test it?
Assurance has to scale with generation, or it becomes a queue.
Can we explain it?
To a reviewer, to an auditor, to the person on call at 3am.
Can we measure it?
Flow, reliability, value, capability. Not lines, not accepted suggestions.
Can we safely reverse it?
Reversibility is the cheapest safeguard you will ever buy.
If the answer is "no," increasing delegated authority is hard to justify. Chapter 1 · Ledger L1–L5
Chapter 3 · the task-shape heuristic
Six dimensions that shape the verification burden of a delegated task.
AI-favourable
Lower expected verification burden
AI-hostile
Higher expected verification burden
Heuristic, not a predictive model.
Use it to shape authority, reviews, and safeguards, then validate it against your own delivery outcomes.
See report: ch. 3 · Ledger: L6–L11
Chapter 7 · the pairing
One organisation sits at one maturity level. Every workflow inside it sits at its own authority band. That is why these are two axes and not one number.
Both marked cells below belong to the same M4 organisation. High maturity does not mean maximum authority everywhere.
Not a scoring surface. No target cell, no diagonal to travel. Unmarked cells carry no judgement.
The one-page version
Executive summary
Every number traces to an entry in the evidence ledger. If a number’s permitted framing does not fit the chart, the number does not go in the chart.
Using an agent and surrendering authority to one are different decisions.
Stack Overflow 2025 reported coding-agent use by ~31% of item respondents. A late-April 2026 pulse (n=1,100, different instrument) reported 59% at some frequency, while 63% rarely or never permitted fully autonomous operation.
There is no universal productivity number.
METR’s RCT — 16 experienced OSS developers, 246 tasks, familiar mature repositories — found participants took 19% longer with AI while believing they had been sped up. A narrow historical counterexample, not a current estimate.
More work can move into the system while verification pressure rises.
In Faros telemetry (~22,000 developers), high-adoption periods showed more completed work per developer and, in the same comparison, median review time ~441% higher. Correlational; not customer value.
More code is not more value.
Generated lines, accepted suggestions, and PR counts are activity metrics. Evaluation belongs on a ladder: activity, flow, reliability, value, capability. No activity metric may quietly imply a value outcome.
The evidence justifies stronger controls, not panic.
Package hallucination and security-benchmark results demonstrate durable failure modes. None of these are production prevalence rates, and no credible industry-wide prevalence measurement was identified.
The junior pathway must evolve, not disappear.
Broad verification capability still depends on implementation fluency. The junior role should be redesigned around supervised delivery: not removed, and not converted into cheap AI-review labour.
What's inside
Method
Inclusion
Primary surveys, peer-reviewed and preprint research, platform telemetry, vendor engineering blogs, named-analyst reports, regulatory filings, documented incidents.
Exclusion
Aggregator and SEO content. We do not infer net industry benefit or harm by counting positive against negative studies.
Positioning
Fieldstate is vendor-independent, not neutral. What we commit to is transparent method, graded evidence, and falsifiable claims.
The governing rule
If a number’s permitted framing does not fit the chart, the number does not go in the chart.
Free · no form · no follow-up sequence
Thirty-nine pages, thirty-seven ledger entries, and a framework you can run against your own delivery system this quarter.
Want it walked through with your team? Get in touch.