Practice 02
AI & Intelligent Systems
Explore AI architectures that connect useful reasoning to controlled action: grounded context, explicit orchestration, permissioned tools and evidence-led evaluation. These are design approaches and potential engagements, scoped to your systems and risk appetite—not claims of completed client delivery.
Follow the architecture
Give intelligence a controlled path to action.
Follow a grounded AI request through retrieval, inference and tool use. Keep authorisation outside the model and human approval around consequential actions.
Request — Identify intent
Authenticate the caller and establish a permitted task before passing context to a model. A user prompt is input, not an authority to expand permissions.
- Apply task and tenant boundaries
- Limit prompt size and sensitive inputs
- State where human review is required
Retrieval — Ground the context
Retrieve relevant material that the caller is entitled to access. Treat retrieved instructions as untrusted content, even when they come from an internal document.
- Filter by source permissions
- Track freshness and provenance
- Test irrelevant and conflicting evidence
Model — Generate a proposal
Use the model to propose an answer or a plan. Evaluate accuracy, refusal behaviour and cost against representative tasks rather than relying on a polished demonstration.
- Set token and latency budgets
- Evaluate grounding and uncertainty
- Keep prompts and model versions traceable
Tools — Enforce permission
A tool gateway validates each requested action independently of the model. Restrict scope, validate arguments and require approval for high-impact operations.
- Allowlist tools and validate schemas
- Use least-privilege credentials
- Require approval before consequential writes
Evaluation — Close the loop
Capture enough evidence to reproduce failures without indiscriminately storing personal information. Re-evaluate when models, tools or retrieval sources change.
- Redact sensitive telemetry
- Monitor quality and failure modes
- Retain a rollback route
Hover, focus or tap a flow node to reveal the decisions behind it.
Architecture, explored
Security is a system. Not a checkpoint.
Design questions to resolve
Start with a boundary. Build the bigger picture.
These are illustrative patterns, not a description of a deployed client system or a guarantee of compliance. The right architecture depends on your constraints, risks and operating model.
From concept to implementation
Technical scope & design decisions
Explore the questions, techniques and potential outputs that can shape an engagement. Final scope depends on the requirements and evidence available.
01Agent orchestration & workflow selection
Start with a deterministic workflow when the steps are known. Introduce model-directed planning only where variable tasks justify the added latency, cost and uncertainty.
Design & implementation focus
- Compare a single model call, a fixed workflow and an agent against the same task set.
- Use manager-led delegation for owned subtasks; use handoffs when control genuinely changes.
- Give every loop a turn limit, deadline, budget and explicit completion condition.
- Parallelise independent tasks only when their writes and dependencies cannot conflict.
Potential engagement outputs
- Pattern decision record
- Task graph and ownership map
- Budget and termination policy
Read the underlying guidance: Anthropic · OpenAI Agents SDK · Microsoft Agent Framework
Discuss this capability →02MCP-connected tools & enterprise actions
Model Context Protocol provides an interface for tools and context. It is not a substitute for identity, business authorisation or reviewing what an external server can access.
Design & implementation focus
- Authenticate the caller and authorise each operation outside the model.
- Validate tool arguments, scope credentials and restrict egress to approved systems.
- Review servers and tool descriptions as a supply-chain boundary; do not trust retrieved instructions.
- Separate read-only discovery from writes and bind approval to the exact action.
Potential engagement outputs
- Tool contract catalogue
- Permission and consent matrix
- Integration conformance tests
Read the underlying guidance: Model Context Protocol · OpenAI Agents SDK
Discuss this capability →03Agent-to-agent interoperability
A2A supports communication between independently implemented agents. Use it when a real organisational or runtime boundary exists, rather than making every function a remote agent.
Design & implementation focus
- Define the remote agent capability, task contract and permitted information exchange.
- Specify authentication, task cancellation, failure responses and artifact validation.
- Carry correlation IDs across handoffs without leaking unrelated conversation history.
- Test version compatibility and timeouts; a protocol does not prove that a remote result is correct.
Potential engagement outputs
- Cross-agent contract
- Trust-boundary diagram
- Interoperability test pack
Read the underlying guidance: Agent2Agent protocol
Discuss this capability →04Retrieval, context & memory engineering
Ground answers in accessible evidence and keep working context distinct from durable memory. More stored conversation is not automatically more useful or appropriate.
Design & implementation focus
- Enforce document and tenant permissions before retrieval results reach the model.
- Track citations, source freshness, conflicting evidence and answer abstention.
- Set retention and deletion rules for memory; avoid saving secrets or unneeded personal information.
- Evaluate retrieval recall separately from answer grounding and context-window cost.
Potential engagement outputs
- Retrieval evaluation set
- Context and retention design
- Evidence provenance model
Read the underlying guidance: OpenAI Agents SDK · Microsoft Agent Framework
Discuss this capability →05Long-running agents & managed execution
Asynchronous work needs a recoverable execution model, not an endless chat request. Compare a managed harness with a self-operated runtime against control, residency and operational requirements.
Design & implementation focus
- Persist task state and resume from a known checkpoint after interruptions.
- Use idempotency keys for side effects and reconcile outcomes before retrying.
- Make cancellation, deadline expiry and human escalation explicit.
- Assess sandbox isolation, network access, session retention and provider-specific beta limitations.
Potential engagement outputs
- Runtime trade-off assessment
- Recovery and cancellation runbook
- Side-effect ledger design
Read the underlying guidance: Claude Managed Agents · Microsoft Agent Framework
Discuss this capability →06Agent evaluations, tracing & release gates
Measure whether the system completes the right task safely, not just whether its final answer sounds convincing. Combine offline regression tests with monitored, reversible rollout.
Design & implementation focus
- Build representative and adversarial task sets with known expected outcomes.
- Score tool selection, argument validity, grounding, approvals and successful completion.
- Trace model, prompt, retrieval and tool versions while redacting sensitive content.
- Set latency and cost thresholds, compare baselines and define rollback triggers before release.
Potential engagement outputs
- Versioned evaluation suite
- Trace and redaction specification
- Release scorecard
Read the underlying guidance: OpenAI Agents SDK · Microsoft Agent Framework
Discuss this capability →07Human oversight & approval journeys
Keep people in control of consequential actions. An approval screen must show what will change, not simply ask whether an agent may continue.
Design & implementation focus
- Show the intended target, payload, impact and rollback options before approval.
- Bind approval to the proposed operation and re-check authorisation when executing it.
- Expire stale approvals and handle rejection without quietly trying another path.
- Provide clear progress, cancellation and handover to a responsible operator.
Potential engagement outputs
- Approval interaction design
- Escalation matrix
- Audit evidence requirements
Read the underlying guidance: OpenAI Agents SDK · Microsoft Agent Framework
Discuss this capability →08Agent security & isolation
Treat prompts, retrieved documents and tool responses as untrusted inputs. Layer validation, least privilege and isolated execution instead of relying on model instructions as a security boundary.
Design & implementation focus
- Separate system policy from data and validate actions at the application boundary.
- Restrict filesystem, network and credential access for code and browser tools.
- Test data-exfiltration attempts, forged tool results and cross-tenant access failures.
- Capture relevant audit evidence, limit retention and maintain a kill switch for tool execution.
Potential engagement outputs
- Agent threat model
- Isolation and access design
- Adversarial test plan
Read the underlying guidance: Model Context Protocol · Claude Managed Agents
Discuss this capability →09AI-native software engineering
Coding agents can assist with repository exploration, tests and change proposals. Keep engineering ownership, review and reproducibility with the delivery team.
Design & implementation focus
- Provide bounded repository context and exclude credentials and confidential datasets.
- Execute agent-proposed changes in isolated workspaces with scoped access.
- Require tests, provenance and human review before merging or deploying.
- Compare assisted delivery against quality, review effort and maintenance outcomes—not lines generated.
Potential engagement outputs
- Assisted engineering workflow
- Repository access policy
- Quality and review scorecard
Read the underlying guidance: OpenAI Agents SDK · Claude Managed Agents
Discuss this capability →10Voice, multimodal & document workflows
Speech, images and documents introduce additional privacy, latency and ambiguity concerns. Scope the interaction and fallback route before choosing model or runtime features.
Design & implementation focus
- Define consent, recording retention and the handling of sensitive attachments.
- Validate extracted fields against schemas and source evidence before downstream actions.
- Design interruption, correction and text alternatives for voice interactions.
- Benchmark end-to-end latency and failure recovery on representative inputs.
Potential engagement outputs
- Interaction and fallback design
- Input evaluation corpus
- Privacy and retention requirements
Read the underlying guidance: OpenAI Agents SDK
Discuss this capability →11Model routing, serving & economics
Select models against measurable task requirements. A cheaper model can increase total cost if retries, poor retrieval or failed tool calls dominate the workflow.
Design & implementation focus
- Compare quality, throughput, latency and total cost per completed task.
- Route by permitted data location and capabilities, not just model price.
- Set concurrency, timeout, context and token limits at the application layer.
- Test fallback behaviour and repeat evaluations when providers or model versions change.
Potential engagement outputs
- Model selection matrix
- Capacity and cost model
- Fallback validation plan
Read the underlying guidance: OpenAI Agents SDK · Microsoft Agent Framework
Discuss this capability →12AI governance & lifecycle assurance
Connect technical controls to named owners and documented purposes. A useful governance model can explain why an agent exists, what it may do and when it must stop.
Design & implementation focus
- Maintain a register of agents, tools, models, owners and approved use cases.
- Assess data protection, licensing and sector obligations with qualified advisers.
- Review changes to prompts, retrieval sources, tools and autonomy as release changes.
- Define incident handling, retirement and deletion processes with observable evidence.
Potential engagement outputs
- AI system register
- Risk and control mapping
- Lifecycle review checklist
Read the underlying guidance: Model Context Protocol · Microsoft Agent Framework
Discuss this capability →Primary references · reviewed 25 October 2026
Grounded in current platform documentation
These references inform architectural options, not endorsements or evidence of MNB client delivery. Features and release status change; verify suitability before adoption.
- Anthropic — workflow and agent design patterns ↗
Foundational pattern guidance; the article notes that tooling has evolved.
- OpenAI Agents SDK — tools, handoffs, sessions, oversight and tracing ↗
Evaluate the runtime and model features against the required environment.
- Model Context Protocol — security best practices ↗
Protocol integration still requires application-level access controls.
- Agent2Agent protocol — interoperability overview ↗
Agent communication is distinct from tool integration.
- Microsoft Agent Framework — agents, workflows and integrations ↗
Feature availability differs by language, provider and release; verify preview status before adoption.
- Claude Managed Agents — managed harnesses and sessions ↗
The reviewed documentation labels this offering beta; assess lifecycle and deployment constraints.