The Model AI Governance Framework for Agentic AI of IMDA (Singapore, January 2026) has 4 dimensions. Each one matches a technique used for agents.
Dimensions
| IMDA dimension | Matching technique |
|---|---|
| 1 Assess and bound the risks upfront | Choose workflow or agent, least privilege, mandates |
| 2 Make humans meaningfully accountable | Approval for risky actions, and logs |
| 3 Implement technical controls and processes | Testing with pass^k, and safeguards in code |
| 4 Enable end-user responsibility | Clear limits, and showing the real total before paying |
Governance is engineering written down, with named people responsible.
Advantages and limitations of agents
| Advantages | Limitations |
|---|---|
| Finish multi-step tasks and not just answer questions | Errors add up over many steps |
| Adapt to what they see along the way | Cost grows quickly with steps and agents |
| Open standards (MCP, A2A, AP2) make tools easy to plug in | Prompt injection is unsolved, and actions raise the stakes |
Summary of the week
| Stage | Key idea |
|---|---|
| 1 What an agent is | An LLM in a loop: decide, act, look, repeat, where the LLM proposes and code acts |
| 2 How LLMs use tools | Fine-tuning teaches tool calls, feedback teaches goals, and MCP plugs tools in |
| 3 Planning, memory, teams | Plan first, check in code, retrieval is memory, and context is a budget |
| 4 Reliability and evaluation | Errors compound as , report pass^k, and check the outcome and the path |
| 5 Agentic commerce | Mandates, tokens and agent identity make paying trustworthy |
| 6 Security and governance | Assume the agent can be tricked, and limit what it can do |
Code
A governance log, so that a person can be held accountable for each step:
import json, time
def audit(step, call, obs, approved_by=None, path="agent_audit.jsonl"):
"""Write one line per step: what was proposed, what came back, and who approved it."""
record = {"time": time.time(), "step": step, "call": call,
"obs": str(obs)[:500], "approved_by": approved_by}
with open(path, "a") as f:
f.write(json.dumps(record) + "\n")Notes:
approved_bynames the person for steps that needed approval, which is the human accountability of dimension 2- the log also settles disputes, as with the signed mandate
The whole week is seen in one running example, see Shopping Agent.