Free GH-600 Exam Flashcards

It is bounded, has clear inputs and outputs, and can be checked against explicit success criteria. Ambiguous ownership or an unverifiable result signals that the step needs further decomposition or human control.

The broad goal with no checkpoints. It hides intermediate decisions and makes drift, unsafe action, and failure diagnosis harder; bounded stages with reviewable artifacts make control possible.

Inputs, outputs, and success criteria. Together they define what the agent may rely on, what it must produce, and how a reviewer or automated check decides whether the work succeeded.

Separation creates a control point: the proposed approach can be inspected and validated before tools change code, infrastructure, or workflow state.

The intended steps, dependencies, affected scope, expected outputs, and validation points. A plan that cannot be inspected cannot support a meaningful pre-execution gate.

To catch scope errors, unsafe assumptions, missing dependencies, and policy conflicts before they become actions. Validation asks whether the plan can achieve the stated criteria within its boundaries.

Place a hard approval gate between planning and execution, with action-capable tools unavailable until the plan passes automated checks or receives the required human authorization.

Autonomy should narrow as operational, security, compliance, or irreversibility risk rises. Higher-risk actions need stronger guardrails, controlled paths, or human approval.

Inspectable artifacts in normal development tools—such as plans, branches, pull requests, logs, checks, and decision records—paired with intervention points for material risk or failure.

Give it only the tools required for the defined task. Adding capabilities that do not serve the success criteria expands the failure and security surface without improving the workflow.

The workflow must also define how the tool is invoked, what inputs and outputs are allowed, which environment it reaches, and how failures or invalid responses are handled.

They should be least-privilege and task-specific: read, write, network, repository, and environment access should be no broader or longer-lived than the task requires.

It exposes defined external capabilities or context through the Model Context Protocol. The agent still needs deliberate configuration and permission boundaries; adding the server does not automatically make every capability appropriate.

A registry supports discovery and approved distribution of MCP servers; an allow list limits which servers may actually be used. Discovery does not itself grant execution permission.

The task's data, tools, credentials, network access, repository state, isolation needs, and environment constraints. The context must enable the task without exposing unrelated resources.

Repository scope limits the code, instructions, secrets, and change surface available to the agent, reducing accidental cross-repository access and making ownership and review clearer.

Its trigger, permissions, inputs, execution environment, produced artifacts, success or failure conditions, and escalation path. CI invocation should be reproducible and auditable rather than an unbounded background action.

It isolates agent changes from the protected branch and routes them through normal checks and pull-request review. Creating a branch or PR is a review boundary, not permission to merge.

Detect and honor the target environment's policies, available tools, credentials, resource limits, and deployment boundaries instead of assuming development, test, and production contexts are interchangeable.

Retry only when the failure is plausibly transient, the operation is safe to repeat, and a bounded retry policy exists. Deterministic, policy, or unknown failures should stop or escalate instead of looping.

Error handling, bounded retries, rollback or recovery, and escalation. Preserve traces and action records throughout so responsibility and the resulting state can be reconstructed.

Use short-term memory for the current interaction, long-term memory for durable reusable facts or preferences, and external memory when state needs independent storage, retrieval, governance, or sharing across sessions and tools.

Store only task-relevant information with a defined purpose and scope. More memory is not automatically better: irrelevant or excessive context increases drift, conflict, privacy risk, and retrieval noise.

They prevent stale or no-longer-relevant context from silently steering future work. Current GitHub Docs describe Copilot Memory as public preview: it validates repository facts against cited code and automatically deletes unused entries after 28 days.

Recorded progress, completed checks, pending steps, decisions and rationale, relevant artifacts, and the last validated environment state. A transcript alone may not express authoritative task state.

Compare current goals, constraints, decisions, repository state, and outputs with the last validated plan and durable artifacts. On divergence, refresh context or stop for review rather than continuing from stale assumptions.

Establish an authoritative state source, compatible representations, freshness checks, access controls, and conflict resolution. Shared state without version or freshness checks can spread stale or contradictory context.

Expected outcomes and operational constraints. A result can be functionally correct yet still fail if it violates security, scope, cost, timing, or workflow constraints.

Quantitative signals are measurable values such as check results, counts, or latency; qualitative signals judge qualities such as clarity, maintainability, or appropriateness. Robust evaluation often needs both.

The criteria were not aligned with development intent. An agent can optimize the metric it is given while missing the real product or engineering goal, so passing must represent the intended outcome—not just proxy checks.

The scanner generates consistent evaluation signals for properties such as test status, code quality, dependency risk, or security findings; the reviewer supplies contextual judgment that automated evidence cannot replace.

Logs, the plan, tool traces, outputs, checks, and workflow artifacts. Correlation helps locate where observed behavior first diverged from intent instead of blaming only the final output.

Reasoning errors, tool misuse, and context or environment issues. The corrective action should target the demonstrated class rather than applying a generic prompt change to every failure.

Revise the agent's instructions, workflow structure, or constraints. Then rerun the same relevant evaluation so the effect of the change is attributable and reviewable.

Refine what is stored and retrieved, narrow scope, improve freshness validation, and adjust expiration, pruning, or reset behavior. Do not use memory tuning to mask a tool or reasoning defect.

Narrow the available tools or permissions, clarify invocation constraints, improve validation and error handling, and require approval for higher-risk operations. Retest against the failure case and normal cases.

Task dependencies, opportunities for safe parallelism, specialization needs, shared-resource conflicts, and how results will be combined and reviewed—not simply the desire to use more agents.

Isolation prevents unintended sharing of context, credentials, files, and partial changes. It also makes each agent's inputs, outputs, failures, and resource use attributable.

Detect the overlap before integration, compare intent and dependencies, serialize or re-scope the work, and reconcile through validation and review. Silently accepting whichever output arrives last loses work and evidence.

Reviewable audit artifacts identifying each producing agent, task and inputs, relevant decisions, actions, outputs, checks, ordering, and handoff status. That evidence lets reviewers reconstruct responsibility and outcome.

The completed work, authoritative artifacts, decisions and rationale, unresolved risks, assumptions, remaining task, expected output, and acceptance criteria. A vague handoff invites duplicated effort and context loss.

It reconstructs how combined agent decisions and interactions produced the final outcome, including delayed effects, repeated patterns, and coordination failures visible only across the full workflow history.

A failed execution terminates unsuccessfully; a partial execution produces only part of the required result; a stalled execution stops making progress without clean completion. Each needs different detection and recovery criteria.

Rollback to a known state, reassign or restart bounded work, and introduce a human-in-the-loop decision. Recovery should preserve evidence and avoid repeating unsafe or conflicting actions.

Version its role and interfaces, complete or transfer in-flight work, preserve state and audit records, validate the replacement, and keep workflow continuity. Retirement should remove execution access without erasing accountability.

Operational impact, security exposure, compliance obligations, reversibility, affected scope, and the need for human judgment. Risk classification determines proportionate autonomy and controls.

Maximize safe delivery speed while meeting organizational security and Responsible AI standards. Too much autonomy creates unmanaged risk; unnecessary approval for low-risk work creates delay without material protection.

Actions requiring contextual judgment or carrying material operational, security, compliance, Responsible AI, or irreversible consequences. The approval point should occur before the consequential action.

A guardrail must block the action, record the attempted path and applicable policy, and route it to an approved alternative or escalation. A warning alone is insufficient when the policy requires prevention.

Least privilege limits what the agent can reach; controlled paths require explicit authorization, protected workflows, and auditable checks for irreversible or compliance-sensitive changes. Together they reduce both opportunity and impact.