From 17 Agentic Patterns to 5 That Survive Compliance

AI agents process routine decisions autonomously. High-risk, low-confidence, or novel decisions route to human reviewers with full context.

This pattern directly satisfies the EU AI Act's human oversight requirements (Article 14). It creates a natural audit trail: every decision is either auto-approved with logged justification, or human-approved with reviewer identity and timestamp.

In practice, the governance layer assigns a confidence score based on data completeness and policy clarity. Above 95%: autonomous execution with full logging. Between 80% and 95%: flagged for asynchronous human review. Below 80%: synchronous human approval required before execution.

Consider a KYC agent processing customer due diligence. Standard cases with complete documentation pass through automatically. Cases with missing documents, politically exposed persons, or unusual patterns escalate to compliance analysts. The agent's preliminary assessment, supporting evidence, and the specific policy clause that triggered escalation all travel with the case.

Regulatory alignment: EU AI Act Article 14, NIST AI RMF GOVERN function, Treasury FS AI RMF control objectives for model governance.

The agent critiques its own reasoning before presenting outputs. Every reflection step is logged in tamper-evident storage.

Research shows reflection can boost performance on coding benchmarks from 80% to 91%. But the compliance value is not the accuracy gain. It is the reasoning chain. When a regulator asks "Why did your system recommend this?", you can show the agent's self-critique, the corrections it made, and the final reasoning. Not just the output.

The cycle works like this: agent generates an initial response, the reflection module evaluates consistency with policy, confidence justification, and edge cases. Each reflection step is logged with timestamp, evaluation criteria, and outcome. The final output includes a provenance chain linking input to initial reasoning to reflection to corrections to final output.

A healthcare example makes this concrete. A clinical decision support agent recommends a treatment plan. Before presenting it, the reflection module checks for conflicts with known allergies, dosage guideline compliance, and drug interactions. Each check is logged, creating a defensible reasoning chain for HIPAA compliance.

Regulatory alignment: EU AI Act Articles 12-13, NIST AI RMF MAP function, HIPAA audit requirements.

Agents invoke external tools through structured, validated interfaces. Every tool call, parameter, and response is logged.

Tool use follows a four-phase cycle: define available tools with structured schemas, let the LLM select and parameterise a tool call, invoke the tool, integrate results. Schema validation makes tool invocations deterministic and auditable. You know exactly which tools the agent called, with what parameters, and what it received back.

In financial services, a portfolio rebalancing agent uses tools to fetch market data, calculate allocations, and execute trades. Each tool call is schema-validated. The allocation tool cannot exceed concentration limits because the schema enforces constraints. Every tool call is logged, creating a complete audit trail from data retrieval through calculation to execution.

Tools with side effects (writes, transactions, communications) require additional governance gates. That distinction matters because it draws a clear line between read operations and actions that change state.

Regulatory alignment: EU AI Act Article 9, SR 11-7, NIST AI RMF MEASURE function.

The agent decomposes complex tasks into structured plans, then executes each step with governance gates between phases.

Planning agents analyse requirements, identify dependencies, and sequence operations before anything runs. The governance checkpoints between steps create natural intervention points. If the plan goes off-course at step 3, you halt execution before steps 4 through 7 cause damage.

How this works: the agent generates a structured execution plan with defined steps, dependencies, and expected outcomes. The plan is logged and, for high-risk operations, reviewed before execution begins. Between each step, a governance checkpoint evaluates whether the plan is still valid. Step outputs are validated against expected outcomes before proceeding. Any deviation triggers re-planning or human escalation.

A clinical trial workflow illustrates the value. The AI system decomposes the process: protocol design, patient screening, data collection, analysis, reporting. At each checkpoint, the governance layer validates compliance with IRB requirements, data privacy rules, and regulatory reporting obligations. The complete chain is logged for FDA submission.

Regulatory alignment: EU AI Act Article 9, NIST AI RMF GOVERN function, Treasury FS AI RMF implementation controls.

This is the pattern nobody writes conference talks about. It is also the one that works.

A single, well-configured agent operating within explicit constraint boundaries: input validation, output filtering, action limits, full logging. For the 70% of use cases where multi-agent orchestration is overkill, a single agent with proper guardrails delivers equivalent results at one-third the cost. The auditability is dramatically simpler.

A customer service agent handles account enquiries. It can retrieve account information and answer questions. It cannot modify accounts, initiate transactions, or access data outside the customer's profile. These constraints are enforced at the infrastructure level, not by the model's judgement. The agent is auditable because its behaviour space is bounded.

Why does this beat multi-agent in regulated settings? Multi-agent orchestration complexity grows near-exponentially when agents delegate to other agents. Coordination overhead becomes the bottleneck. Race conditions and cascading failures are hard to reproduce. A single agent with guardrails has a predictable, bounded behaviour space that compliance teams can validate and regulators can audit.

This pattern aligns with every framework listed because it is inherently the simplest to document, validate, and maintain compliance for.