SEARCH THE COMMONS
Find the mechanism, module, or role.
208 entries · local, instant, no tracking · press / or Cmd+K anywhere on the site
How to Design an Agentic System
Create the first Atlas design canvas and risk profile.
→ Course 002 · FoundationsShould This Be an Agent?
Write an agent-versus-workflow decision memo for Atlas.
→ Course 002 · FoundationsThe Agent Loop and Its Control Boundaries
Implement a bounded tool-calling loop with typed state and stop conditions.
→ Course 002 · FoundationsTool Contracts, Side Effects, and the Acknowledgment Gap
Add typed, risk-rated tools with idempotency and error semantics.
→ Course 002 · Context & DataContext Engineering, Retrieval, and Grounding
Build an ACL-aware context assembler with provenance and citations.
→ Course 002 · Context & DataState, Sessions, and Memory
Separate run state, thread state, and governed long-term memory.
→ Course 002 · OrchestrationOrchestration Patterns and State Machines
Convert Atlas into an explicit graph with routing, parallel work, and join conditions.
→ Course 002 · RuntimeDurable and Long-Running Agents
Make Atlas survive crashes, waits, retries, and deploys without repeating unsafe effects.
→ Course 002 · RuntimeHuman-in-the-Loop and Bounded Authority
Add a durable approval queue with approve, edit, reject, and escalation paths.
→ Course 002 · InteroperabilityMulti-Agent Design, MCP, and A2A
Expose one Atlas capability through MCP and evaluate whether an A2A boundary is justified.
→ Course 002 · TrustSecurity, Identity, Isolation, and Governance
Threat-model Atlas and add policy enforcement, tenant isolation, and adversarial tests.
→ Course 002 · QualityAgent Evaluation and Testing
Create a regression dataset and score components, trajectories, outcomes, and safety.
→ Course 002 · OperationsObservability, SLOs, and Incident Response
Instrument Atlas traces, metrics, logs, dashboards, and an incident replay workflow.
→ Course 002 · OperationsLatency, Cost, Throughput, and Scale
Load-test Atlas and introduce routing, caching, queues, backpressure, and cost budgets.
→ Course 002 · OperationsDeployment, Change Management, and Enterprise Readiness
Ship a versioned canary release with rollback, data controls, and operational ownership.
→ Course 002 · InterviewThe 45-Minute Agentic System-Design Interview
Deliver and self-score a complete Atlas system-design answer.
→ Course 002 · CapstoneCapstone: Atlas in Production
Build, demo, evaluate, and defend the complete production system.
→ Course 001What Is a Long-Running Agent?
Understand what makes long-running agent execution fundamentally different from an ordinary LLM request.
→ Course 001Build the Naive Agent
Build something that works, in plain Python, before introducing any framework.
→ Course 001State: What Must Survive?
Separate conversation, execution state, artifacts, and external reality, and make state represent evidence, not claims.
→ Course 001Durable Execution, Checkpoints, and Replay
Make the agent survive application failure using LangGraph with a PostgreSQL checkpointer.
→ Course 001Idempotency, Side Effects, and Reconciliation
Learn the concept most agent tutorials skip: a checkpoint does not guarantee an external operation happened exactly once.
→ Course 001Work Ownership: Claims, Leases, Heartbeats, and Orphans
Guarantee that exactly one worker executes a run at a time, and that a run whose worker dies is reclaimed automatically.
→ Course 001Failure Handling, Retries, Timeouts, and Degradation
Stop treating every exception as retry(3). Build a failure taxonomy and route each class differently.
→ Course 001Human-in-the-Loop and Durable Waiting
Show how a workflow can stop for hours without keeping a worker occupied, and resume with full state.
→ Course 001Context Engineering Across Long Runs
Treat execution state and model context as separate engineering surfaces; keep the model's working set small, evidenced, and reconstructible.
→ Course 001Observability and Run UX
Make every run explainable to a human in under a minute, from the dashboard, from the trace, and from the database.
→ Course 001Security: The Hostile Vendor
Bound what untrusted content can make the agent do: tool permissions, SSRF controls, prompt-injection boundaries, and tenant isolation.
→ Course 001Evaluating Long-Running Agents
Make evals the development loop: error analysis on real traces, binary criteria, a validated judge, and a regression gate that proves version B beats version A.
→ Course 001The Deep Agents Bridge
Rebuild the agent on a higher-level harness, verify every invariant still holds, and learn to tell framework features from guarantees.
→ Course 001Industry Architectures and Case Studies
Recognise the same six primitives you built inside every credible production long-running agent, and spot the marketing that hides their absence.
→ Course 001Deploying and Operating the Agent
Put the agent on a public URL, for free, and prove that a redeploy mid-run does not lose work.
→ Course 001Production Failure Gauntlet
Run every failure at once against your deployed agent, measure it, and publish a reliability report anyone can reproduce.
→ PracticeCloud Security Investigation and Remediation Agent
Practice scenario, 45 minutes. Focus: authority, tool safety, durable execution, incident response.
→ PracticeEnterprise Coding Agent for a Million-Line Repository
Practice scenario, 45 minutes. Focus: repository context, sandboxing, long-running execution, evaluation.
→ PracticeInsurance Claims Intake and Adjudication Support Agent
Practice scenario, 45 minutes. Focus: documents, human authority, long-running workflows, governance.
→ PracticeCustomer Support Agent with Refund Authority
Practice scenario, 45 minutes. Focus: interactive latency, action authority, scale, fraud controls.
→ PracticeEnterprise Research Agent with Document ACLs
Practice scenario, 45 minutes. Focus: retrieval, access control, citations, memory.
→ PracticeLong-Running Procurement and Vendor Coordination Agent
Practice scenario, 45 minutes. Focus: long-running workflows, external communication, approvals, multi-agent boundaries.
→ PracticeHealthcare Prior-Authorization Preparation Agent
Practice scenario, 45 minutes. Focus: sensitive data, evidence and policy, urgent workflows, human authority.
→ PracticeProduction Incident-Response Copilot
Practice scenario, 45 minutes. Focus: real-time operations, reversible actions, observability, human coordination.
→ PracticeFinancial Research and Trade-Proposal Agent
Practice scenario, 45 minutes. Focus: freshness, evidence, market-data licensing, action separation.
→ PracticeAutonomous Data Analyst over an Enterprise Warehouse
Practice scenario, 45 minutes. Focus: code execution, data governance, semantic layer, cost control.
→ PracticeTravel Disruption Recovery Agent
Practice scenario, 45 minutes. Focus: burst traffic, external transactions, user preferences, multi-step compensation.
→ PracticeEnterprise Agent Platform for Multiple Product Teams
Practice scenario, 45 minutes. Focus: platform boundaries, multi-tenancy, governance, developer experience.
→ Reference · MethodThe RUNTIME design method
The seven-step method the whole course applies to every system.
→ Reference · CanvasAgentic System Design Canvas
Twelve sections that turn a vague agent idea into a defensible design.
→ Reference · ReferenceFailure taxonomy
The failure classes an agentic system must route deliberately.
→ Reference · TemplateTool contract template
The typed, risk-rated contract every tool needs before it ships.
→ Reference · TemplateThreat model template
Assets, threats, boundaries, prevention, detection.
→ Reference · TemplateEval plan template
What to measure, on which data, with which release gate, across components, trajectories, and outcomes.
→ Reference · TemplateSLO and incident runbook
Indicators, objectives, alerts, and the runbook an operator uses at 3am.
→ Reference · TemplateArchitecture decision record
One decision, its options, the choice, and what it costs.
→ Reference · CapstoneAtlas reference implementation spec
The full specification of the canonical course system.
→ Reference · CapstoneCapstone brief
What to build, submit, and defend for the Atlas capstone.
→ Reference · ReferenceConcept coverage matrix
Every concept in the course, mapped to the module that teaches it.
→ Reference · AssessmentCourse rubric
How a finished capstone is judged, dimension by dimension.
→ Reference · AssessmentInterview scorecard
The scorecard to self-score a 45-minute design interview.
→ Reference · AssessmentQuiz answer key
Every quiz answer with the reasoning behind it, so a wrong pick becomes a lesson rather than a guess.
→ Reference · ReferencePrimary source map
Every implementation-sensitive claim, mapped to its official source.
→ Reference · ReferenceGlossary
Production-oriented definitions for the whole vocabulary.
→ ConceptCheckpoint
A durable snapshot from which an interrupted run can continue.
→ ConceptReplay
Reconstruct execution from recorded state and decisions.
→ ConceptDurable execution
Progress that survives process and infrastructure failure.
→ ConceptIdempotency
Repeating an operation has the same intended effect as doing it once.
→ ConceptReconciliation
Compare desired and observed state before deciding what to do next.
→ ConceptLease
Time-limited ownership that must be renewed while an owner is healthy.
→ ConceptHeartbeat
A periodic signal that proves a worker is still healthy.
→ ConceptFencing token
A monotonic token that prevents stale owners from writing.
→ ConceptOrphaned run
Work whose recorded owner is no longer able to make progress.
→ ConceptHuman interrupt
A durable pause that hands a decision to a person.
→ ConceptContext compaction
Reduce working context while retaining decision-relevant evidence.
→ ConceptArtifact
A durable, inspectable output passed between steps or people.
→ ConceptTrajectory evaluation
Judge the path an agent took, not only its final answer.
→ ConceptEvaluation judge
A rubric-driven evaluator for behavior that cannot be checked exactly.
→ ConceptAgent harness
The control layer around a model: tools, state, policy, budgets, and verification.
→ PatternReconcile Before Retry
Observe the external system before repeating an ambiguous side effect.
→ PatternEvidence-Carrying Claim
Attach inspectable evidence to every material agent assertion.
→ PatternBounded Agent Loop
Give every loop explicit step, time, and cost limits.
→ PatternDurable Human Wait
Persist state before pausing for an asynchronous human decision.
→ PatternContext Quarantine
Keep untrusted context isolated until it is inspected and promoted.
→ PatternArtifact Handoff
Pass durable artifacts, not hidden conversational state, between actors.
→ PatternVerify Then Act
Check authority, preconditions, and evidence before an irreversible action.
→ PatternState Outside the Model
Keep authoritative workflow state in a durable system of record.
→ PatternBudget Envelope
Bound cost, time, tokens, and tool calls for every run.
→ PatternLease and Recover
Use expiring claims and fencing to safely reclaim abandoned work.
→ Career segmentFrontier Model Deployment
Turn model capability into a production customer workflow.
→ Career segmentAgent / Application FDE
Integrate an AI application into a real business workflow.
→ Career segmentEnterprise Data / Platform FDE
Adapt a platform to customer data, applications, and operations.
→ Career segmentInference / ML Systems FDE
Make models fast, reliable, scalable, and economical in production.
→ Career segmentEvaluation / Observability FDE
Make AI quality measurable in development and production.
→ Career segmentVertical AI FDE
Combine engineering depth with a regulated or specialized domain.
→ Career segmentPublic Sector / Mission FDE
Deploy into constrained, secure, and mission-critical environments.
→ Career segmentPhysical AI / Robotics FDE
Deploy software, models, and hardware into real environments.
→ PathSoftware / AI Engineer → FDE
Add discovery, production judgment, field communication, and evidence to your engineering base.
→ PathSolutions Architect / SE → FDE
Build coding depth, failure reasoning, stateful systems, evaluation, and reproducible proof.
→ PathExisting FDE → Senior FDE
Practice architecture judgment, portfolio patterns, executive communication, and turning field learning into product improvements.
→ SpecializationAgent / Application FDE
Integrate AI applications with tools, state, context, evals, security, and operational ownership.
→ SpecializationEnterprise / Platform FDE
Adapt platforms to customer data, identity, applications, infrastructure, and operating constraints.
→ SpecializationInference / ML Systems FDE
Make models fast, reliable, scalable, observable, and economical in production.
→ GlossaryA2A (Agent2Agent Protocol)
A protocol for agents or agent services to discover capabilities, exchange tasks, communicate progress, and return artifacts across system or organizational boundaries. It does not replace local authorization or policy.
→ GlossaryAcknowledgment gap
The interval after an external effect may have occurred but before the caller durably records success. Retrying blindly during this gap can duplicate the effect.
→ GlossaryAction
A model-proposed or workflow-selected next step. An action is not executed until the application validates and authorizes it.
→ GlossaryActivity
In a durable workflow system, a unit of nondeterministic or external work such as a model call, API request, file operation, or database effect. Activities may execute more than once and therefore need effect-safe semantics.
→ GlossaryAgent
A software system in which a model dynamically selects among allowed actions based on observations and state to pursue a goal. The surrounding application owns authority, persistence, and enforcement.
→ GlossaryAgent loop
The bounded sequence of assemble context → request decision → validate → authorize → execute → observe → update state → stop or repeat.
→ GlossaryAgentic workflow
A workflow that includes one or more model-controlled decisions but retains explicit application-owned control flow.
→ GlossaryArtifact
A versioned, persistent output or large input such as a due-diligence packet, document, prompt snapshot, code patch, or report.
→ GlossaryAuthority boundary
The line between what the model may propose and what the application, policy engine, or human may authorize.
→ GlossaryBackoff
Increasing delay between retries, usually with jitter, to avoid amplifying a dependency failure.
→ GlossaryBackpressure
A mechanism that tells upstream producers to slow down, defer, or reject work when downstream capacity is saturated.
→ GlossaryBounded autonomy
Agent behavior constrained by allowed actions, scope, state preconditions, budgets, stop rules, and approval policy.
→ GlossaryBulkhead
Resource isolation that prevents one tenant, workload, dependency, or tool from exhausting the entire system.
→ GlossaryCache key
The complete identity of reusable data. In an agent system it may need tenant, authorization, source version, freshness, model, prompt, parser, and policy dimensions.
→ GlossaryCanary release
A production rollout to a small, representative subset of live traffic before broader deployment.
→ GlossaryCapability
A tool, resource, prompt, or agent service that can perform a bounded function. Discovery of a capability does not grant permission to use it.
→ GlossaryCheckpoint
Persisted execution state from which a run or graph can resume. A checkpoint is not necessarily the same as a durable workflow event history.
→ GlossaryCircuit breaker
A dependency-protection pattern that temporarily stops calls after failures exceed a threshold, then probes recovery.
→ GlossaryClaim-level provenance
A mapping from each material generated claim to one or more source identifiers and exact locators.
→ GlossaryCompensation
A business action that reverses or mitigates a completed external effect. It is not the same as rolling back an in-process database transaction.
→ GlossaryContext assembly
Selecting, filtering, ranking, formatting, and budgeting the instructions, state, evidence, history, and tool information provided for one model decision.
→ GlossaryContext engineering
Designing what information and capabilities are available to a model at each step, including selection, compaction, provenance, trust, and token budgets.
→ GlossaryControl plane
The deterministic software layer that validates, authorizes, schedules, persists, observes, and recovers agent behavior.
→ GlossaryCritical path
The longest dependency chain that determines minimum end-to-end latency.
→ GlossaryDead letter
Work that cannot complete after defined handling and is moved to an explicit queue or state for investigation rather than retried forever.
→ GlossaryDecision schema
A typed structure that limits model output to allowed fields and actions.
→ GlossaryDeterministic replay
Reconstructing workflow state by re-running control-flow code against recorded events while producing the same decisions for the same history.
→ GlossaryDurable execution
Persisting enough execution history or state for work to resume after crashes, deploys, or long waits without restarting unsafe effects.
→ GlossaryEffect
A change visible outside the local decision process, such as sending a message, charging a card, publishing an artifact, or modifying a system of record.
→ GlossaryEffect reconciliation
Querying the authoritative external system to determine whether an uncertain effect occurred before deciding to retry, compensate, or escalate.
→ GlossaryError budget
The amount of unreliability permitted by an SLO over a window. Some security requirements remain zero-tolerance and are not ordinary error budgets.
→ GlossaryEval case
A versioned input fixture, expected properties, rubric, and metadata used to measure system behavior.
→ GlossaryEvaluator-optimizer
A pattern in which one step generates work and another evaluates it against explicit criteria, repeating within a bound.
→ GlossaryEvent history
An ordered durable record of workflow events used to restore and audit execution.
→ GlossaryExcessive agency
Granting an agent more capability, permission, scope, or autonomy than required for the user outcome.
→ GlossaryFallback
A validated alternate model, provider, tool, source, or workflow path used when the primary path is unavailable or unsuitable.
→ GlossaryFan-out/fan-in
Running independent branches in parallel and joining them according to a defined completion policy.
→ GlossaryFeature flag
A trusted runtime control that enables, disables, or changes a capability without requiring the model to decide its own configuration.
→ GlossaryFreshness policy
Rules defining how recent evidence or cached results must be for a task or risk tier.
→ GlossaryGolden set
A stable, reviewed collection of representative eval cases used for regression testing.
→ GlossaryGuardrail
A check that can block, modify, or route input, output, or tool behavior. A guardrail is one layer, not a replacement for authorization or isolation.
→ GlossaryHandoff
A control transfer in which another agent becomes responsible for continuing the interaction or task. It differs from calling another agent as a bounded tool and receiving a result.
→ GlossaryHeartbeat
A progress signal from a long-running activity that lets the runtime detect failure and optionally retain a progress checkpoint.
→ GlossaryHITL (Human-in-the-Loop)
A designed workflow state in which an authenticated human reviews, edits, approves, rejects, or redirects work before a controlled transition.
→ GlossaryHuman-on-the-loop
Human oversight of an operating system without reviewing every action, often through monitoring, exception handling, and policy.
→ GlossaryIdempotency
The property that repeating an operation with the same logical request identity produces no additional unintended effect.
→ GlossaryIdempotency key
A stable identifier used by the caller and ideally the external system to deduplicate a logical effect.
→ GlossaryIndirect prompt injection
Malicious or conflicting instructions embedded in retrieved webpages, documents, emails, tool output, or other content rather than directly entered by the user.
→ GlossaryInformation gain
The expected reduction in uncertainty from an action. It can help prioritize research under step, time, and cost budgets.
→ GlossaryInstruction/data separation
Treating system/developer policy as instructions and retrieved/user content as untrusted data that cannot redefine authority.
→ GlossaryJoin policy
Rules for when parallel branches may proceed: mandatory results, optional results, timeouts, degraded states, and escalation.
→ GlossaryJitter
Random variation added to retry delays so many clients do not retry simultaneously.
→ GlossaryLeast privilege
Granting each user, service, tool, and workflow only the permissions needed for the current task and scope.
→ GlossaryLong-term memory
Governed information retained across threads or tasks, such as a durable user preference or organization fact. It is not ordinary run state.
→ GlossaryMCP (Model Context Protocol)
A protocol for a host to connect to servers that expose resources, prompts, and tools through negotiated capabilities. The host remains responsible for trust, authorization, consent, and data handling.
→ GlossaryMemory poisoning
Introducing false, malicious, or inappropriate information into retained memory so later behavior is influenced.
→ GlossaryModel gateway
A centralized layer for model access, approved aliases, routing, credentials, quotas, structured output, telemetry, and fallback.
→ GlossaryModel judge
A model used to grade another model or agent output. It requires a clear rubric and calibration against human labels.
→ GlossaryMulti-agent system
A system with multiple independent or semi-independent model-driven components. Separate agents should be justified by trust, lifecycle, ownership, specialization, or scaling boundaries, not character names.
→ GlossaryNo-progress loop
Repeated decisions or actions that do not materially change evidence, state, or outcome.
→ GlossaryNon-deterministic work
Work whose result can vary or depends on the external world, such as time, network calls, model output, or random values. Durable runtimes record it outside replayable control code.
→ GlossaryObservation
A normalized, persisted result made available to the workflow or agent after tool execution, policy, or user input.
→ GlossaryOrchestrator
The component that coordinates steps, state, tools, and agents. It may be deterministic, model-assisted, or mixed.
→ GlossaryOutcome eval
Evaluation of the final user or business result rather than only an intermediate model output.
→ GlossaryPolicy engine
A trusted component that evaluates identity, state, risk, resource, and action rules outside natural-language model instructions.
→ GlossaryPrompt cache
Reuse of a stable prompt prefix or provider-side computation. It is an optimization, not a state or correctness mechanism.
→ GlossaryPrompt injection
Input that attempts to change or override intended instructions, reveal data, or induce unauthorized behavior.
→ GlossaryProvenance
Metadata showing where information came from, when it was retrieved, which version was used, and how it supports an output.
→ GlossaryQuality gate
A release condition that must pass before a new model, prompt, workflow, policy, or tool version receives more traffic.
→ GlossaryRate limit
A provider or application constraint on requests, tokens, concurrency, or work over time.
→ GlossaryReason code
A controlled categorical explanation used for routing, audit, or evaluation. It is easier to measure than free-form rationale.
→ GlossaryRelease bundle
The versioned combination of code, workflow, prompts, model policy, retrieval configuration, tools, policy, schemas, and datasets that defines production behavior.
→ GlossaryReplay-safe code
Workflow control code that produces the same decisions when reconstructed from its recorded history.
→ GlossaryRetrieval grounding
Supplying relevant external evidence to a model and constraining outputs to evidence-backed claims.
→ GlossaryRetry budget
A cap on additional work caused by retries, protecting the system from amplification.
→ GlossaryRisk tier
A classification that determines allowed autonomy, required evidence, model route, review, retention, and operational controls.
→ GlossaryRPO (Recovery Point Objective)
Maximum acceptable data loss measured in time or committed work.
→ GlossaryRTO (Recovery Time Objective)
Maximum acceptable time to restore a service or capability.
→ GlossaryRun
One business execution of an agentic process, potentially spanning many services and long waits.
→ GlossarySandbox
An isolated execution environment with bounded filesystem, network, process, resource, and credential access.
→ GlossarySemantic cache
Reuse based on input similarity rather than exact identity. It requires careful scoping and is risky for time-sensitive or consequential outputs.
→ GlossaryService-level indicator (SLI)
A measured property of service behavior, such as valid completion rate or P95 latency.
→ GlossaryService-level objective (SLO)
A target for an SLI over a time window.
→ GlossarySession state
Conversation- or thread-scoped state used across several turns. It should have explicit ownership, retention, and compaction.
→ GlossaryShadow execution
Running a new version against mirrored or captured input without allowing production effects.
→ GlossarySide effect
See **Effect**.
→ GlossarySignal
An external event delivered durably to a waiting workflow, such as approval, cancellation, or document arrival.
→ GlossaryState machine
An explicit set of states and allowed transitions with guards and effects.
→ GlossaryStep budget
Maximum agent decisions or actions allowed in a run.
→ GlossaryStructured output
Model output constrained or validated against a schema.
→ GlossarySupervisor pattern
One orchestrator delegates bounded work to specialists and remains responsible for global state and policy.
→ GlossaryTenant isolation
Preventing one customer or business unit from accessing or affecting another across identity, storage, retrieval, caches, queues, tools, memory, telemetry, and operations.
→ GlossaryThread state
State retained for a conversation or task thread. Often synonymous with session state, but define the exact scope.
→ GlossaryTimer
A durable scheduled wake-up used for retries, expiry, reminders, and escalation without holding a process thread.
→ GlossaryTool
A typed capability available to the agent through an application-controlled gateway.
→ GlossaryTool gateway
A trusted component that validates, authorizes, executes, observes, rate-limits, and audits tool calls.
→ GlossaryTool-selection eval
Evaluation of whether the correct tool, order, arguments, and error handling were chosen.
→ GlossaryTrace
A causal record connecting work across services and asynchronous steps for one execution.
→ GlossaryTrace grading
Evaluating the sequence of model, tool, policy, state, and approval events rather than only the final output.
→ GlossaryTrajectory
The ordered series of decisions, actions, observations, and transitions taken during a run.
→ GlossaryTrust boundary
A boundary across which data or authority changes trust level and therefore requires validation, authentication, authorization, or isolation.
→ GlossaryUncertain write
An external write whose completion status is unknown because the response was lost or the caller crashed. It requires reconciliation.
→ GlossaryUntrusted content
User, retrieved, uploaded, or tool-generated content that may be false, malicious, or instruction-bearing. It should not inherit system authority.
→ GlossaryValid completion
A run that not only finishes, but also satisfies required evidence, policy, schema, authority, and safety conditions.
→ GlossaryVector store
A storage/indexing system that supports similarity search over embeddings. It is one possible retrieval component, not a memory or authorization system by itself.
→ GlossaryWorkflow
Application-owned control flow with explicit steps and transitions. A workflow can include model decisions without giving the model full control.
→ GlossaryWorkflow versioning
Techniques that preserve compatibility for in-flight durable executions while code and behavior evolve.
→ GlossaryZero-tolerance gate
A release or runtime condition for which any occurrence is unacceptable, such as cross-tenant leakage or prohibited autonomous publication.
→Nothing matches that. Try a shorter term.