Pattern 01
Bounded single-agent execution
Let one agent choose its next observation while code owns permissions, budgets and stopping conditions.
Build an incident-response assistant by making one system contract explicit at a time. This collection contains 15 design patterns, 45 articles, and 15 runnable Python examples. Every pattern has a design article explaining requirements and tradeoffs, a build article tied to the actual source, and an operate article covering failure drills, security, observability, cost, and acceptance.
Read patterns 01–05 to choose execution and coordination boundaries. Continue through 06–10 for events, durable state, scoped memory, approval, and delegated tool authority. Patterns 11–15 connect repeat-safe effects, tenant isolation, bounded model routing, controlled releases, and regression discovery. Follow a complete three-article pattern when solving a specific problem, or use the numbered order to assemble the broader architecture.
The downloadable examples use Python 3.11 or later and the standard library. They are deterministic local exercises with synthetic data and explicit adapter seams. They do not call a model, authenticate a real user, deploy cloud resources, or perform production remediation. Storage and failure behavior vary by pattern; each walkthrough states its exact boundary. Windows execution was checked. macOS commands are usage guidance and have not been verified on a physical Mac.
Architecture and sequence diagrams are original conceptual illustrations by Sandeep Belgavi. Build diagrams map the local teaching behavior; design and operating diagrams also describe proposed extensions. A passing fixture suite supports the illustrated contract, not a claim that the complete incident platform is production-ready. All workload numbers, route costs, and release thresholds are explicitly illustrative.
Start with the Evals collection for evidence, coverage, precision, and release-measurement fundamentals. The Harness Engineering collection provides complementary context, and the labs collection offers further interactive practice. Source references accompany the specific external claims; the incident scenarios, code, and architectural recommendations are original educational material.
Download all 15 Python examples and their tests · Run instructions and scope · Evals prerequisites
01 / Execution and control
Pattern 02
Deterministic workflows with model-driven steps
Keep the incident procedure fixed while a constrained model supplies classifications or summaries at explicit stages.
Pattern 03
Planner-executor with independent verification
Accept bounded plans, execute observation steps and assess their claims against a separately supplied evidence view.
02 / Coordination
Pattern 04
Supervisor and specialist routing
Route a scoped incident to a known specialist while keeping authorization, fallback and result validation under supervisor control.
Pattern 05
Parallel workers with controlled aggregation
Run independent observations concurrently, preserve failures and disagreement, and aggregate under a contract that does not mistake speed for truth.
Pattern 06
Event-driven agent collaboration
Coordinate incident workers through durable events, explicit ownership and replay-safe local state.
03 / State and knowledge
Pattern 07
Durable execution and recovery
Preserve incident workflow progress across restarts without confusing a checkpoint with an external transaction.
Pattern 08
Layered memory and evidence retrieval
Retrieve incident context by tenant, revision and freshness while preserving citations and deletion boundaries.
Pattern 09
Human approval and safe resume
Bind approval to a specific proposal, reviewer, revision and expiry before consuming it for a local dispatch marker.
04 / Trust and effects
Pattern 10
Secure tool gateways and delegated identity
Enforce issuer, audience, actor, tenant, scope and target boundaries before an incident agent reaches a provider adapter.
Pattern 11
Idempotent actions and compensation
Stable operation identity, durable receipts, and explicit reversal for incident-response effects.
Pattern 12
Tenant-isolated agent services
Bind identity, evidence, caches and resource accounting to an authorized tenant context.
05 / Production operations
Pattern 13
Budget-aware model routing and overload control
Admit only eligible work, reserve resources before dispatch, and reconcile usage explicitly.
Pattern 14
Versioned deployments, canaries and rollback
Treat release inputs as a bundle and separate quality decisions from authorized infrastructure changes.
Pattern 15
Ambient quality and controlled improvement
Resolve evidence, discover regressions, review fixtures, and retain an independent release boundary.