Articles

Article · 5 min

From the artefact to the tool: controlled autonomy in software development

Traditional, AI-assisted, agentic - the ways of working coexist. What shifts: people work less on the code and more on the system that produces it.

Editorial team ·

Traditional, AI-assisted and agentic development exist side by side today. The difference is not how much the AI does, but what is built around it.

Four ways of working

  • Ad-hoc use - AI answers questions alongside the work. Helpful, but it changes nothing.
  • Vibe coding - software emerges in rapid conversation with little structure. Good for exploring, not a delivery discipline.
  • AI-assisted development - AI contributes inside the engineer’s workflow; the human stays in the execution loop. The everyday mode.
  • Agentic development - engineers delegate bounded outcomes with explicit context, acceptance criteria and automated verification. Humans keep risk-appropriate checkpoints.
Engineering framelowhighDegree of delegationhighlow
Vibe codingIntent becomes software through rapid conversation, with little explicit engineering structure.Useful for exploration. Not, by itself, a delivery discipline.
Agentic developmentEngineers delegate bounded outcomes with explicit context, constraints, acceptance criteria and automated verification.Humans keep risk-appropriate checkpoints and accountability.
Ad-hoc useAI answers questions alongside the work, without a real handover and without a frame around it.Useful for a question. It does not change how the work gets done.
AI-assisted developmentAI contributes inside the engineer's workflow; the engineer stays continuously in the execution loop.The everyday mode. Verification happens continuously, inside the work.
illustrativeOutputTimeHigh engineering frameLow engineering frame
Engineering frame
specification, context, testing and verification around the handover
Degree of delegation
how much is handed over before a human looks at it
Four ways of working: they differ less in how much the AI does than in what surrounds it · “Vibe coding” after Karpathy, 2025

What makes agentic development professional is the engineering frame: specification, context, tests, verification. The stronger it is, the more the AI can carry.

From the artefact to the tool

In traditional development, people write code. In agentic development, they build and maintain the agent system - skills, tool connections, rules - that writes the code. Our principle: when an agent errs, we correct not only the workpiece but the machine.

Traditional
Personwrites code
↓
IDEdevelopment environment
↓
Artefact in the repository
AI-assisted
Personprompts, reviews suggestions
↓
IDE + assistantsuggestions in the editor
↓
Artefacts in the repository
Agentic
Personbuilds and maintains the agent system
↓
Agent systemskills · tools · rules · agents
↓
Many artefacts - reviewed before acceptance
Human work moves from the artefact to the tool - accountability and approval stay

How much the AI carries per phase

less more
  1. Requirements & analysisAI roleAnalystWhat the AI contributesConsistency checks, user stories, acceptance criteriaAI share of execution0+AssistedHuman work vs the same phase without AI-0+Human roleIntentCheckpoint✓ Acceptance criteria agreed
  2. Design & architectureAI roleArchitecture peerWhat the AI contributesSolution options, dependency and impact analysisAI share of execution0+AssistedHuman work vs the same phase without AI-0+Human roleDecisionCheckpoint✓ Architecture decision approved
  3. ImplementationAI roleCo-developerWhat the AI contributesCode creation, refactoring, larger coherent changesAI share of execution0+Controlled autonomyHuman work vs the same phase without AI-0+Human roleDirectionCheckpoint✓ Code review before merge
  4. Testing & QAAI roleTest engineerWhat the AI contributesTest generation, coverage, quality and security analysisAI share of execution0+DelegatedHuman work vs the same phase without AI-0+Human roleAssuranceCheckpoint✓ Quality gates passed
  5. DeploymentAI roleRelease assistantWhat the AI contributesBuild and deploy configuration, pipeline error analysisAI share of execution0+DelegatedHuman work vs the same phase without AI-0+Human roleApprovalCheckpoint✓ Release approval
  6. Operations & monitoringAI roleDiagnosticianWhat the AI contributesIncident and root-cause analysis, log evaluationAI share of execution0+AssistedHuman work vs the same phase without AI-0+Human roleInterpretationCheckpoint✓ Findings assessed before action
↻ Each cycle's learning starts the next one
The engineering frameWays of working · specifications · context upkeep · test and environment · infrastructure · guardrails · tooling
AI share of execution0+Controlled autonomyHuman work vs working without this frame-0+
From assistance to controlled autonomy: how much the agent carries differs by phase - what stays with our engineers changes in kind, not in amount

Human effort does not disappear. It moves - out of execution, into verification and into the frame.

Accountability cannot be delegated

AI-generated changes pass the same review and release process as any other change. Each phase ends with a decision a person is accountable for. And client data is protected by architecture and contract, not goodwill: no model training, no data retention at the provider, processing in the EU, only the context the task needs.

ConcernGuarantee
Code used to train models
→
✓ No model training - contractually excluded
Data logged or retained
→
✓ Zero data retention - nothing persists after processing
Dependence on one provider
→
✓ Several vetted providers with equivalent guarantees
Unclear rights in results
→
✓ Licence terms checked before use
Processing outside the EU
→
✓ Processing in the EU on approved paths
Tools adopted without assessment
→
✓ Assessed and approved before project use
Client data and IP - protected by architecture and contract, not goodwill

See it in practice in our project Test operations with agents and our practice experiment From ticket to pull request.

Part of the SII Germany team

Let’s
talk.