Agentic AI Engineering
Agentic AI Engineering
Designing, Evaluating, and Operating Systems That Act
Komal Nakrani
First edition/Agentic AI Engineering · Volume 1
Agentic AI Engineering
Designing, Evaluating, and Operating Systems That Act
A professional field book for converting justified model-directed autonomy into bounded, inspectable, evaluable, recoverable action across tools, state, people, and production systems.
- Edition
- First edition
- Version
- 1.0.0
- Published
Table of contents
01The Profession Behind the AgentDefine Agentic AI Engineering by responsibility for bounded action and separate model, harness, product, platform, domain, and authority decisions.65 min02Earn the Right to Use AutonomyCompare fixed, assisted, single-agent, and multi-agent execution against one task and choose the least complex adequate level.70 min03Contract the Goal, Actions, and AuthorityTurn a vague delegated request into a machine-checkable contract for completion, actions, effects, budgets, approval, stop, escalation, and evidence.80 min04Make the Loop InspectableTranslate the task contract into explicit run states, events, deterministic gates, cancellation, replay, and externally verified completion.85 min05Treat Tools as Capabilities, Not FunctionsDesign narrow typed capabilities with legible intent, effect classes, deterministic validation, explicit errors, idempotency, reconciliation, and test doubles.85 min06Bind Identity, Delegation, and ConsentKeep human, service, agent, task, resource, delegation, approval, and effect identities distinct while enforcing audience, scope, expiry, revocation, and audit.85 min07Engineer State, Context, Artifacts, and MemorySeparate authoritative state, finite model context, scratch work, durable artifacts, preferences, and memory through provenance, permission, freshness, isolation, retention, and deletion.90 min08Start With One AgentIntegrate the accepted contracts into one reproducible FieldOps vertical slice and freeze a measurable single-agent baseline before topology experiments.85 min09Add Handoffs and Multiple Agents Only With EvidenceTreat routing, specialists, workers, reviewers, and remote agents as controlled topology hypotheses with explicit handoffs, ownership, authority, cancellation, aggregation, cost, and rejection evidence.90 min10Make Execution Durable and Effects RecoverablePreserve run ownership and effect correctness across crashes, retries, duplicate events, lost responses, leases, deployments, reconciliation, and compensation.95 min11Design Human Checkpoints That Can Actually WorkMake human checkpoints durable, timely, evidence-rich, authority-bound, rejectable, expiring, cancellable, resumable, and capable of explicit takeover.90 min12Build a Representative Task EnvironmentVersion a synthetic world of state, users, tools, policies, budgets, faults, consequences, splits, and omissions so evaluation claims remain bounded and reproducible.90 min13Evaluate Outcomes, Actions, State, and EfficiencyBuild claim-specific evidence across outcomes, trajectories, state, effects, authority, recovery, efficiency, calibrated judgment, and consequence segments.115 min14Attack Assumptions and Inject FailureTurn trust, identity, tool, memory, approval, runtime, and effect assumptions into executable containment tests with independent controls and named residual owners.120 min15Trace Runs Without Turning Logs Into SurveillanceCorrelate model, tool, state, approval, control, and effect transitions through minimal causal records, sealed evidence, bounded access, and deletion-aware retention.100 min16Budget Quality, Time, Cost, and CapacityOperate agent workloads inside consequence-aware joint envelopes for quality, latency, actions, cost, queues, concurrency, and dependency capacity.100 min17Release, Interrupt, Recover, and LearnIncrease exposure only to answer a named release question, keep stop authority reachable, and turn incidents into verified changes to code, tests, contracts, or controls.120 min18Interoperate Without Surrendering SemanticsPlace MCP tools and A2A remote agents behind pinned adapters that verify discovery and preserve local identity, authority, effect, timeout, cancellation, and evidence semantics.110 min19Replay Change Across Model, Tool, Policy, and HarnessTreat every behavior-affecting component change as an evidence question, replay outcomes and trajectories, migrate state, bound rollout, and retire superseded claims.115 min20Lead the Evidence, Not the HypeDefend the complete evidence chain, test patterns across cases, choose retain, reduce, repair, reuse, platformize, or retire, and hand mature capabilities to named owners without hiding uncertainty.110 min