Agentic Engineering
The agent is not the engineering system.
Coding agents analyse, plan, change, test and open pull requests. They only become an engineering system once context, identity, tools, permissions, verification and delivery are designed together.
Five planes
What gets built around the agent
Control
Policy, identities, permissions, audit and cost. Who may do what is written down as a rule and checked by a machine.
Context
AGENTS.md, architecture, ADRs, specifications. The agent receives a bounded, versioned bundle instead of a mountain of knowledge.
Execution
Isolated workspace, its own identity, approved tools. No developer credentials, and the network is closed by default.
Verification
One command checks everything: build, tests, types, SAST, SCA, architecture rules. Green tests alone are not enough.
Delivery
Branch, pull request as evidence, independent CI, human review, merge. A person stays accountable.
The path of one task
Issue → task contract → policy → sandbox → agent → tools → verification → pull request with evidence → CI → review → merge
The path
From finding to operation
Readiness
Where does your organisation stand? One repository, one team, read against the requirements of the standard. The result is a maturity level, the gaps and a 90-day path.
Target architecture
Gateway, context service, policy engine, sandbox, tool gateway, verification, telemetry: which components you need and which of your existing products can carry them.
Pilot in 90 days
Three phases, each with an acceptance gate before an agent is allowed more. Separate paths for existing systems and for new ones.
The foundation
Agentic Engineering Standard 2026
Behind the assessment stands a standard of my own, vendor-neutral: 48 requirements, five planes, autonomy levels A0 to A5, maturity levels 0 to 5, a 90-day model, JSON schemas, Rego policies and a runnable reference repository. It is public, as draft version 1.1.
Implementation patterns
What of it is already built
The instruments are older than the standard. It sums up what is applied in them and in my own systems. It is not a certification.
| Instrument | What it implements |
|---|---|
| Local Agent Pipeline | An approved action space, a human-in-the-loop gate, a tamper-evident audit trail. |
| Agentic AI Governance Toolkit | Risk scoring with control intensities, an action authority matrix, governance findings as a build artefact. |
| AI Deployment Ordinance | Work areas in four levels: free, with a cross-check, proposal only, blocked. |
| RAG Approval Blueprint | Approval file and control catalogue with test procedures: evidence first, approval second. |
Built systems
The same rules on my own systems
StudyWithMe
Separate containers, an API service that cannot read the content, every AI call logged, AI results marked as low-trust, the audit log append-only.
razbiram.com
A live platform, built with a multi-agent workflow of parallel streams, loops and skills.
What it is not
Nothing here sells you a coding agent or a tool, and there is no seal to earn. You decide which agent and which products you use. The standard sets out what each of them has to meet.
Where does your organisation stand? The entry point is the assessment, case B. To the assessment