Agentic AI Systems

Guide · Enterprise AI Solution EngineeringPage 8 of 18Overview → … → Agentic AI

Executive summary

Agentic systems plan, select tools, execute and update state toward an objective. Design for minimum necessary autonomy: tool allowlists, durable state, human approval on high-impact actions, and loop/cost controls—prefer RAG Q&A until tool use is justified.

When to use this page

• Use when multi-step tool use is required beyond grounded Q&A.
• Use after a simpler pattern has been ruled out in AI patterns.
• Defer multi-agent designs until single-agent tool use is stable and observable.

Executive view

Autonomy is a risk decision. Approve write tools and irreversible actions only with clear owners and approval gates.

Decision required: What is the minimum autonomy that still creates value?

Technical view

Include planner, tool registry, policy layer, state store, guardrails, human approval and observability.

Test loop limits, duplicate-action controls, recovery and tenant isolation before production write access.

Owners

Owner · AI Solution Engineer / agentic platform lead

Contributors · Security, product owner, domain SME, SRE, AI engineering, responsible AI

What makes a system agentic?

An agentic system can interpret an objective, plan actions, select tools, execute steps, read results, update state, request approval and continue until completion.

Agent architecture

A production agent typically includes planner, tool registry, policy layer, state store, memory, execution engine, guardrails, human approval, observability and evaluation.

Single-agent versus multi-agent

Use a single agent when the workflow is relatively simple, centralised reasoning is sufficient and operational control is important.

Use multiple agents where roles are clearly separable, specialist context improves performance, parallel execution creates value, or different tools or policies apply.

Do not create multiple agents only because the design appears innovative.

Human-in-the-loop controls

Human approval may be required before sending external communications, updating financial systems, approving payments, changing customer records, making regulated decisions, deleting data or executing high-impact actions.

Agent state

State may include current objective, completed steps, pending actions, tool results, approval status, errors and retry count.

State should be durable, auditable, secure, recoverable and tenant-isolated.

Agent failure modes

Common risks: infinite loops, tool misuse, incorrect planning, prompt injection, excessive cost, duplicate actions, unauthorised access, state corruption and inconsistent recovery.

Agent design principle

Give an agent the minimum autonomy required to create value.

Decision required

Justify autonomy against a simpler pattern. Irreversible tools proceed only with human approval in V1; multi-agent complexity waits until single-agent reliability is proven.

Case study: financial-services assistant

Default mode is not fully agentic. Start with RAG Q&A. Introduce tools only for authorised CRM lookups. Any customer-record update requires human approval. Multi-agent designs are deferred until single-agent tool use is stable and observable.

Output

Autonomy justification, tool allowlist with permissions, approval map by impact, state store and recovery design, and loop/cost/duplicate-action controls.

Common failure modes

• Agents with write access and weak allowlists
• No budget or loop limits
• State in chat history only
• Multi-agent complexity before single-agent reliability
• Missing approval gates on irreversible actions

Solution Engineer checklist

Solution Engineer checklist

• Autonomy justified against a simpler pattern
• Tool allowlist and permissions documented
• Approval points mapped to impact
• State store and recovery tested
• Loop, cost and duplicate-action controls in place

Practical exercise

List every tool an agent might call. Label each read-only, reversible write, or irreversible. Only irreversible tools may proceed with human approval in V1.

Key takeaways

• Minimum autonomy: do not add agents for novelty.
• Durable, auditable state—not chat history alone.
• Human approval for high-impact and irreversible actions.
• Control loops, cost and duplicate actions before scale.

What to do next

1. Continue with Security and privacy.
2. Align oversight with Responsible AI and governance.
3. Instrument agents in Evaluation and observability.
4. Follow the AI Agents roadmap and full production guide.