Project 5 — Tool-calling agent

Requirements

Read claim status, lookup policy, retrieve instructions, ask for missing IDs, and escalate. Start read-only.

Architecture

User → bounded decision loop → validated/authorized tools → observations → answer. Tools are existing business-service wrappers.

Technologies

Provider SDK or chosen agent SDK, Zod tool schemas, existing Node services, run tracking.

Database design

Runs, call/result references, task outcomes, and conversation ownership; no need to duplicate canonical claims.

API design

Agent message/run endpoints; tools called internally through narrowly scoped contracts.

AI design

Small allowed toolset, clear descriptions, typed results, maximum steps/deadline/token budget, and stop/clarify behavior.

Security

Tool allowlist, object authorization, no arbitrary database/URL tool, output minimization, and injected-result tests.

Evaluation

Correct tool and arguments, denied cross-customer IDs, absent data, transient failures, loops, and truthful task completion.

Observability

Per-tool spans, loop steps, error reasons, unnecessary calls, outcome, and cost distribution.

Deployment

Apply provider and service quotas; isolate read-only credentials; keep safe template fallback.

Interview questions

• “Does the model execute the function?” Answer: it proposes; application code authorizes and executes.
• “How do you stop loops?” Answer: explicit progress, step, time, and token checks.

Acceptance: No fabricated status and no unauthorized object access, even under adversarial prompts.