Agentic Dev Benchmark 2026

How the assessment works

Here you can see what we look at in the assessment and which modules it consists of. Every team is different: we always agree the process and focus with you individually. The basic structure stays the same.

What we look at

Maturity model: 9 dimensions

We assess your setup along nine dimensions, from strategy and governance to the team. Dimension 9 measures whether your setup keeps up with the weekly pace of innovation.

The 9 dimensions of the maturity model
No.DimensionKey questionEvidenceKPI / tracking
1 Strategy and governanceAre there goals, owners and an AI usage policy?Policies, roles, budgetsExistence, definition and maturity
2 Agent setup and harnessDo agents work with rules, context and guardrails, or only with prompts?Rule files, hooks, skills, orchestration, model routingOrchestration, routing
3 SpecificationHow good are requirements before an agent starts?Tickets, specs, definition of ready, acceptance criteriaDefinition of ready, quality of acceptance criteria
4 Quality and reviewHow is agent output checked before it is merged?Review process, reviewer ≠ author, coverage, mutation scoreMaker/checker principle, mutation score
5 Security and supply chainHow are secrets, dependencies and risky actions controlled?Approvals, credential isolation, egress, dependency monitoring, SBOMEgress control, credential isolation
6 Compliance and evidenceCan you show what an agent did and when?Audit trail, GDPR, EU AI Act, CRA, BSI C5Quality markers for compliance
7 Delivery and costDoes AI make delivery faster, and what does it cost?Lead time, deployment frequency, token cost per taskCost per task
8 Team and rolesHave developers moved from coder to orchestrator (TAPO)?Skills, enablement, role profilesTAPO concept
9 Evolvability of the harnessWho maintains the setup, how quickly do new practices arrive, is it tested?Owner, versioning, update cadence, regression tests of the rules, drift between teamsDoes the setup keep up with the weekly pace of innovation?

Scale and evidence status

The 5 maturity levels per dimension
LevelMeaning
1 Ad hoc Individual prompts, everyone works differently
2 Repeatable Rule files per repo, not team-wide
3 Defined Team-wide rules, gates manual
4 Enforced Hooks and gates automatic, versioned
5 Measured and learning Metrics, rules regression-tested, ongoing maintenance with an owner

Every checkpoint gets an evidence status: evidenced (artefact seen), open (statement in the interview only) or not checked. A dimension only counts as reliably assessed if at least two thirds of its checkpoints are evidenced. Otherwise the score is marked “preliminary”.

How it works

Package and process

The assessment consists of nine modules. Part A is vendor-neutral and the same for every team. Part B is optional and only follows if you want it. Together, the modules correspond to around 7.5 person-days.

The modules of the assessment
No.ModuleWhat happensPerson-days (estimated)
1 Scoping and pre-assessmentA short questionnaire in advance on all nine dimensions and the key facts of your setup0.5
2 Remote pre-analysisRead-only analysis on your machines with the BeeKeep3r Harness: test coverage, code quality, dependencies and the quality of 3–5 real tickets1.0
3 On-site dayOne day with your team, including a block on the lifecycle of your harness: who maintains it, and how quickly do new practices arrive?2.0
4 Evaluation part AYour report: a vendor-neutral evaluation of all nine dimensions, compact at around 10 pages2.0
5 Part B: comparison with BeeKeep3r + make-or-buyOptional: principles heatmap, make-or-buy calculation with your figures and a plan for a trial0.5
6 Executive debrief60 minutes of results from part A, then an optional 30 minutes on part B0.5
7 Anthropic question channelWe pass your questions on to Anthropic. An answer is not guaranteed. Shared only anonymised and with your written approval0.25
8 Cohort benchmark and roundtableAnonymised comparison with the other participating teams, in ranges only, and a joint roundtable0.25
9 Follow-up after 6–8 weeksTogether we look at what has been implemented and, if you like, discuss a trial0.5
Total 7.5
Agentic Dev Benchmark 2026

Sounds like a fit? Apply now.

First award: 16 Nov 2026 · Application deadline: 30 Nov 2026 · Duration: 3 minutes