Agentic Development Platforms
AI agents don't just write code; they need infrastructure to run it. Solid internal developer platforms are the building blocks of agentic ones, and without them organisations fall into one of three traps: blocked, ungoverned or bottlenecked.
🌱 A rough idea, planted recently. Expect rough edges — it will grow.
One line from chapter 2 of Domain-Driven Platform Engineering stuck with me: agentic AI hasn’t reduced the need for platform engineering, it has increased it. Solid internal developer platforms are the building blocks of agentic development platforms: platforms that support agents doing real work.
Agents need infrastructure too
An agent can finish the code before you finish your coffee. Then it needs a database, a queue, an environment. Who decides what it gets? Humanitec frames the answer as three traps most organisations fall into:
| Trap | What happens |
|---|---|
| Blocked | Agents can’t provision anything. The infrastructure waits on a human ticket, and the AI speed-up disappears one ticket at a time. |
| Ungoverned | Agents provision whatever works: admin roles, unencrypted storage, a database ten times bigger than needed, and no audit trail. Risk and cost grow with every deployment. |
| Bottlenecked | Agents can ask, but a person approves every request. Five agents today, fifty next quarter: the review queue grows with adoption. The bottleneck moved; it didn’t go away. |
The punchline: you adopted AI to move faster, and your platform team has never been slower.
The way out: governed autonomy for agents
The answer is the same one that works for people (The Platform Relevance Gap): self-service within guardrails.
- People set policy once, as code: sizes, encryption, least-privilege roles, budgets.
- The platform enforces it on every request automatically, and records an audit trail.
- Domain defaults come built in, so an agent working on a regulated service gets compliant infrastructure without having to know the rules.
In plane terms, humans work in the control plane and the platform enforces the rules in the data plane. Nobody sits in the request path approving tickets (Designing a Platform for AI Assistants). It also means the checks around agents are deterministic, not another model’s opinion (Don’t Read the Code? Verifying AI-Written Code).
Sources: Humanitec · Domain-Driven Platform Engineering, chapter 2
- #platform-engineering
- #ai-agents
- #governance
- #self-service
- #policy-as-code
↩ Linked from (1)
…ny of it is always-on: scoped tokens, no production credentials, cost caps on agent runs, and an audit trail (Agentic Development Platforms).
↪ Links to (4)
Related by topic
- Growing
AI Before the LLM Boom: Notes from 2019
AI, machine learning - Growing
What Makes an AI Agent
Platform engineering - Growing
Choosing Frameworks and Standards for 80 Sites
Frontend craft - Growing
DORA Metrics
Platform engineering