How Do You Choose an Agent Architecture?

Choose the least complex architecture that satisfies capability isolation, context isolation, policy boundaries, latency, and ownership. Begin with one agent plus narrow tools. Add deterministic workflow steps for known sequences, routing for distinct intents, and specialists only when they need materially different instructions, context, tools, models, or approval rules.

What will you be able to do?

You will compare common patterns, select a SupportOps baseline, and record the conditions under which the design should evolve.

What is the decision map?

Pattern	Use when	Avoid when
Single agent	one outcome and tool surface	prompt has conflicting roles/policies
Deterministic workflow	sequence and branches are known	path depends on open-ended discovery
Router + specialists	intents need different ownership	categories overlap heavily
Manager + agents-as-tools	one agent must synthesize	specialists should own user conversation
Handoff	specialist should take control	caller must retain final responsibility
Parallel workers	independent bounded subtasks	shared mutation or ordering matters
Event-driven agents	long-running, resumable work	simple synchronous request fits

What architecture should SupportOps start with?

Start with one triage agent and two read-only tools: get_ticket and search_knowledge. Return a structured decision. A deterministic application layer applies policy and routes high-risk cases to review.

Add a billing specialist only when it needs a separate knowledge source and approval policy. Add a repository specialist only when file/tool context would pollute normal support work. Keep mutation behind a deterministic proposal/approval/commit workflow.

Rendering diagram...

How do you justify another agent?

Require at least one measurable reason:

smaller context improves quality;
separate tools reduce authority exposure;
different model/cost profile improves an SLO;
trace ownership becomes clearer;
separate team owns the specialist contract; or
independent work reduces latency without introducing mutation conflicts.

“It feels more agentic” is not a reason.

How do you prove the architecture is better?

Run the same task suite against the simpler and proposed designs. Compare task success, unsafe tool calls, handoff accuracy, latency distribution, token/cost proxy, and operational complexity. Include cases that do not require the specialist.

Failure injection: Make two specialists capable of the same task with vague descriptions. Observe routing instability. Fix the capability contracts before tuning the router prompt.

What mistakes should you avoid?

Using agents for deterministic sequencing.
Sharing a broad tool registry with every specialist.
Letting multiple workers mutate the same resource concurrently.
Assuming parallel calls reduce end-to-end cost.
Ignoring who owns the final user-visible answer.

Check your understanding

When should a specialist take ownership rather than act as a tool?
What evidence justifies splitting one agent?
Why is shared mutation dangerous in parallel work?

Expert extension

Write two architecture decision records for SupportOps: the single-agent baseline and a proposed specialist design. Define the evaluation threshold required to approve the split.

Official references

How Do You Choose an Agent Architecture?

What is the decision map?

Pattern	Use when	Avoid when
Single agent	one outcome and tool surface	prompt has conflicting roles/policies
Deterministic workflow	sequence and branches are known	path depends on open-ended discovery
Router + specialists	intents need different ownership	categories overlap heavily
Manager + agents-as-tools	one agent must synthesize	specialists should own user conversation
Handoff	specialist should take control	caller must retain final responsibility
Parallel workers	independent bounded subtasks	shared mutation or ordering matters
Event-driven agents	long-running, resumable work	simple synchronous request fits

What architecture should SupportOps start with?

Rendering diagram...

How do you justify another agent?

Require at least one measurable reason:

smaller context improves quality;

separate tools reduce authority exposure;

different model/cost profile improves an SLO;

trace ownership becomes clearer;

separate team owns the specialist contract; or

independent work reduces latency without introducing mutation conflicts.

“It feels more agentic” is not a reason.

How do you prove the architecture is better?

Failure injection: Make two specialists capable of the same task with vague descriptions. Observe routing instability. Fix the capability contracts before tuning the router prompt.