---
title: What should developers learn before building AI agents?
url: https://opencontroller.lyzr.com/blog/ai-agent-governance-training-for-developers
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# What should developers learn before building AI agents?

**Before building production AI agents, developers should learn six things: how to register an agent and state its scope, how agent identity and delegation work, how to scope tools to least privilege, how prompt injection and tool misuse happen, how to evaluate agents before and after release, and what to do in an incident.** Most of this is ordinary secure engineering applied to software that decides its own next step.

## Why do agent builders need governance training at all?

A developer who builds a web service decides every call it makes. An agent builder decides which calls are possible, and the model picks among them at runtime. That shifts risk from code paths to permissions, tool design and context. Training that only covers prompt writing leaves the riskiest decisions to chance.

There is also a regulatory reason. Article 4 of the EU AI Act, applicable since 2 February 2025, requires providers and deployers of AI systems to take measures to ensure a sufficient level of AI literacy among staff dealing with AI systems. A documented curriculum with completion records is straightforward evidence for that.

## What should the curriculum cover?

| Module | What developers should be able to do | Time |
|---|---|---|
| 1. Registration and scope | Register an agent with owner, purpose, data, tools and risk tier; write a scope statement | 30 min |
| 2. Identity and delegation | Choose between workload identity and on-behalf-of delegation; avoid shared service accounts and long-lived keys | 60 min |
| 3. Tool and MCP scoping | Give each agent the minimum tools; review MCP servers before use; separate read and write tools | 60 min |
| 4. Threats | Explain direct and indirect prompt injection, tool poisoning, memory poisoning and excessive agency; map to the OWASP Top 10 for Agentic Applications | 90 min |
| 5. Evaluation | Build a test set, run multi-turn and red-team evals, set pass thresholds, re-run after every prompt or model change | 90 min |
| 6. Runtime and cost | Route model calls through the approved gateway; tag traffic by agent and version; set budgets | 30 min |
| 7. Logging and evidence | Emit traces with agent and user identity; know what must be retained | 30 min |
| 8. Incident duties | Recognise an incident, escalate, support quarantine, help reconstruct events, retire the agent safely | 30 min |

Roughly seven hours in total. Split it into short sessions with hands-on labs rather than one long course.

## Which threats should every agent builder recognise?

The OWASP Top 10 for Agentic Applications, published in December 2025, is the most useful shared vocabulary. At minimum, builders should be able to explain and give an example of:

- **Agent goal hijack** through injected instructions in documents, emails or web pages.
- **Tool misuse**, where an agent uses a legitimate tool in a harmful way.
- **Identity and privilege abuse**, such as an agent acting with more rights than its user.
- **Memory and context poisoning** that persists across sessions.
- **Cascading failures** across chained agents.
- **Rogue agents** whose behaviour drifts from their intended purpose.

A good lab: give developers a working agent with an email-reading tool and a planted instruction in an inbox, and ask them to make it fail safely.

## What checklist should a developer pass before first production release?

1. Agent registered with owner, scope statement and risk tier.
2. No shared or personal credentials; identity method documented.
3. Every tool justified against the scope statement; write tools listed separately.
4. MCP servers reviewed and pinned to approved versions.
5. Model calls go through the approved gateway with agent and version tags.
6. Eval suite exists, passes thresholds and runs in CI.
7. At least one red-team scenario for indirect prompt injection.
8. Traces include agent ID, version and user identity.
9. Owner knows how to request quarantine and who can pull the kill switch.
10. Retirement plan: how credentials get revoked when the agent is switched off.

## How do you keep training current?

Agent tooling changes monthly. Review the curriculum each quarter against new incidents in your estate, changes to MCP and identity specifications, and updates to OWASP and MITRE ATLAS. Use your own near-misses as case studies; they land better than generic examples.

## How Opencontroller does this

Training sticks when the platform enforces the same rules. Opencontroller gives builders a governed registry for agents, versions, tools and MCP servers; delegated on-behalf-of identity and agent workload identity federation; the LLM Gateway for routing and cost attribution per agent and version; and multi-turn and red-team evals. Opencontroller Edge extends this to developer devices, coding agents, skills and local MCP servers. See [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

Related: [AI literacy for agent users under the EU AI Act](/blog/eu-ai-act-ai-literacy-for-agent-users) and [a production-readiness gate for AI agents](/blog/production-readiness-gate-for-ai-agents).

## Frequently asked questions

**Is prompt engineering training enough for agent builders?**
No. The riskiest decisions in an agent are identity, tool scope and evaluation, not prompt wording.

**Does the EU AI Act require developer training?**
Article 4 requires measures for sufficient AI literacy among staff dealing with AI systems, applicable since 2 February 2025. A documented curriculum is evidence for that.

**How long should agent governance training take?**
Around seven hours across eight short modules, with hands-on labs and a pre-release checklist.

## Key facts

A developer curriculum for AI agent governance covers registration and scope, identity and delegation, tool and MCP scoping, agentic threats (OWASP Top 10 for Agentic Applications, December 2025), evaluation, gateway routing and cost, logging, and incident duties, about seven hours. EU AI Act Article 4 (AI literacy) applies from 2 February 2025. Opencontroller provides a governed registry, on-behalf-of identity, LLM Gateway, evals, and Opencontroller Edge for developer devices; it starts from $250K per deployment per year.

## Related

- [Does the EU AI Act still require AI literacy for people who use AI agents?](https://opencontroller.lyzr.com/blog/eu-ai-act-ai-literacy-for-agent-users.md)
- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
