How long does the assessment take?

About 8–10 minutes. Twenty-five multiple-choice questions across five sub-dimensions of agentic AI.

What does the score mean?

Your total maps to one of five readiness bands: Nascent → Exploring → Operational → Advanced → Leading. Each band comes with prioritized advice for the next investment to make.

What are the five sub-dimensions?

Tool use & integration, Memory & state, Planning & orchestration, Guardrails & safety, Evaluation & observability. Each gets 5 questions and a per-dimension score so you can see where you are strong and where you are weakest.

Where does this fit in the GenAI Maturity Framework?

This quiz is the operational counterpart to the Agentic AI dimension of the framework (see /framework/dimensions/agentic). Agentic capability becomes material at framework maturity level 4-5.

Are my answers stored anywhere?

No. Nothing is submitted or logged. The tool runs entirely in your browser; your answers persist to localStorage on your device only, so you can come back later.

GenAI MaturityGAI

Framework Try Assessments Use Cases Industries Implementation Learning Paths Blog Resources

Home
›
Tools
›
Agentic AI Readiness

Agentic AI Readiness Score

A free 25-question assessment of your organization's ability to run agentic AI in production, across five sub-dimensions: tools, memory, planning, guardrails, and observability. About 10 minutes. Nothing is submitted.

25 questions · max score 75 · operational counterpart to the Agentic AI framework dimension

Progress

0 / 25

Tool use & integration

How agents call external tools and services — from ad-hoc integrations to standardized protocols like MCP.

1. How do your agents discover and call external tools?
No agents in production yet.Hard-coded integrations per agent, one adapter per tool.Shared internal library / registry — but bespoke per agent.Standardized protocol (Model Context Protocol or equivalent) with dynamic discovery.
2. How do you restrict which tools an agent can call at runtime?
No enforcement — agent can call anything wired in.Documented in prompt but not enforced.Per-agent allowlist enforced in orchestration layer.Per-agent, per-user, per-context allowlists with runtime policy engine.
3. How are tool outputs sanitized before feeding back to the model?
Not sanitized — raw tool output flows back into context.Basic length limits only.Structured schema validation; string outputs sanitized.Full untrusted-input treatment — schema validation + injection scanning + provenance labels.
4. How do agents authenticate when calling tools that require credentials?
Shared service account — no per-user attribution.Long-lived API keys per agent.Short-lived per-agent credentials via token exchange.Per-user OAuth flow with least-privilege scopes; MCP with OAuth 2.1 or equivalent.
5. How do you handle destructive / irreversible tool actions (send email, transfer funds, delete data)?
Agent can execute directly — no confirmation.Post-hoc log review; no gate.Human confirmation for a defined set of high-risk actions.Policy-driven gate — every action classified for reversibility; irreversible actions always human-confirmed.

Memory & state

How agents maintain and retrieve state across turns and sessions — from stateless to persistent shared memory.

1. How do agents maintain context across steps within a single task?
Stateless — every call is independent.In-memory conversation buffer within one process.Persistent per-task state store with checkpointing.First-class state machine with checkpoints, replay, and step-level rollback.
2. How do agents remember information across sessions or users?
No cross-session memory.Ad-hoc storage in application DB.Dedicated memory service — per-user, retention rules defined.Governed memory tier — access controls, deletion propagation, per-record classification.
3. How does the agent retrieve relevant context from long-term memory?
Does not.Keyword search over a store.Vector similarity search over embeddings.Hybrid retrieval (vector + BM25 + graph) with re-ranking and provenance.
4. How is agent memory audited and cleared?
Not audited.Manual deletion on request.Documented retention policy; automatic expiry.Automated retention + data-subject-request propagation + audit log of memory reads.
5. How does memory handle cross-agent knowledge sharing?
No sharing — each agent has its own silo.Shared read-only knowledge base, no write-back.Shared read/write, but no isolation between agents or tenants.Tenant-scoped shared memory with role-based access control and change history.

Planning & orchestration

How agents decompose goals into steps — from linear chains to reasoning-model-driven dynamic plans.

1. How do agents decompose complex goals into steps?
Single-shot prompts, no decomposition.Fixed prompt chains authored by developers.Model-generated plans at start of task; static thereafter.Dynamic replanning via reasoning models with observation → revise loop.
2. How do you choose which model handles which subtask?
Single model for everything.Hand-coded routing rules per task type.Cost/quality routing — cheap model for simple, premium for complex.Automated router with quality feedback loop; reasoning models for planning-heavy subtasks.
3. How do agents recover from tool failures mid-task?
Fail the whole task.Naive retry (with same input).Retry with backoff + fallback tool.LLM-driven diagnosis + strategy shift + escalation to human when confidence drops.
4. How do multi-agent systems coordinate?
No multi-agent systems.Ad-hoc — one agent calls another as a tool.Framework-based (LangGraph / CrewAI / AutoGen) with defined roles.Protocol-based coordination (A2A or equivalent) with capability advertising and negotiation.
5. How do you cap per-task compute and cost?
No caps.Global rate limit only.Per-agent max steps and max tokens.Per-task budgets across steps/tokens/tool-calls/wall-clock with graceful degradation.

Guardrails & safety

Runtime safety controls — input filtering, output validation, per-action budgets, human-in-the-loop.

1. How do you filter inputs (user + tool + retrieval) before they reach the agent?
No filtering.Basic keyword blocklists.Classifier model for topic, PII, and jailbreak attempts.Layered filters — classifier + regex + provenance labels + tenant-specific policies.
2. How do you filter outputs before showing them to users or invoking tools?
No filtering.Basic toxicity check.PII redaction + toxicity + harm-category classification.Layered output check + per-tool schema validation + policy engine.
3. How do you defend against indirect prompt injection (adversarial content in retrieved documents / tool outputs)?
Not addressed.Documented risk, no active control.Retrieval content is delimited and instructions from it are heuristically ignored.Structural separation (channel-typed inputs) + injection classifier + human review on suspect actions.
4. How is the human-in-the-loop pathway implemented?
No HITL.Post-hoc log review.Sync checkpoint on defined action classes with UI approval.Confidence-triggered HITL — auto-decides high confidence, escalates low confidence, with SLO-backed queue.
5. How do you red-team agents before launch?
No red-teaming.Ad-hoc developer testing.Structured red-team suite (prompt injection, jailbreak, misuse) run per release.Automated adversarial harness + independent red team + go/no-go gate on critical findings.

Evaluation & observability

Instrumentation and evaluation of agent behaviour distinct from LLM evals.

1. What do you log for every agent run?
Nothing — no persistent logs.Basic input/output pair.Full trajectory (prompts, tool calls, tool results, retries, cost, latency).Full trajectory + structured metadata + guardrail-trigger events + user feedback linkage.
2. What metrics do you track for agent quality?
None.Task-success rate only.Task-success + trajectory quality + cost-to-solution.Full agent-eval stack — success, trajectory quality, cost, groundedness, guardrail-trigger rate — with SLOs.
3. Do you run offline evaluations on a golden dataset?
No.Occasional manual runs.Regular runs pre-release.CI-gated evals — no ship without pass; dataset expands from real incidents.
4. How do you detect drift in production agent behaviour?
We do not.Manual sampling.Live eval harness compares against golden set at fixed cadence.Streaming eval on sampled prod traffic + automated regression alerts + rollback trigger.
5. How do you close the feedback loop from incidents back into evals and guardrails?
We do not.Postmortems document lessons; no eval/guardrail change.Postmortems mandate at least one eval or guardrail addition.Structured incident-to-eval pipeline — every P1/P2 automatically drafts an eval entry for review.

FAQs

How long does the assessment take?
About 8–10 minutes. Twenty-five multiple-choice questions across five sub-dimensions of agentic AI.
What does the score mean?
Your total maps to one of five readiness bands: Nascent → Exploring → Operational → Advanced → Leading. Each band comes with prioritized advice for the next investment to make.
What are the five sub-dimensions?
Tool use & integration, Memory & state, Planning & orchestration, Guardrails & safety, Evaluation & observability. Each gets 5 questions and a per-dimension score so you can see where you are strong and where you are weakest.
Where does this fit in the GenAI Maturity Framework?
This quiz is the operational counterpart to the Agentic AI dimension of the framework (see /framework/dimensions/agentic). Agentic capability becomes material at framework maturity level 4-5.
Are my answers stored anywhere?
No. Nothing is submitted or logged. The tool runs entirely in your browser; your answers persist to localStorage on your device only, so you can come back later.

Related on this site

Framework dimensions

Agentic AI

Free tools

LLM Model Comparison
AI Compliance Mapper

Whitepapers

Agentic AI Reference Architecture 2026

Comparisons

Agentic AI vs Traditional AI: What Actually Changes
MCP vs REST API for AI Tool Calling: When to Use Which
LangGraph vs CrewAI vs AutoGen: Agent Framework Comparison

Next steps

LLM Evaluation & Assurance— Metrics, benchmarks, red-team
LLM Evaluation Playbook— Six-phase implementation guide

For developers

Call this framework and its tools from your own agent via the Model Context Protocol (MCP) server. Works with Claude Desktop, Cursor, Zed, Continue, and the OpenAI Agents SDK.