ImageFirm / INTELLIGENCE
FIELD GUIDE 02 · OCT 2026
The language of agentic AI

Know the terms.
See the system.

12 essential concepts for understanding how AI agents work—and where their limits belong.

A corrected, expanded edition of your nine-term blueprint. Clear definitions, practical examples and the distinctions that matter.

Explore the terms ↓

A shared vocabulary for better decisions.

12 terms

01Foundation

Harness

The environment that lets a model work.

The runtime coordinates model calls, context, tools, execution state, permissions and recovery. The action loop is one part of that environment.

Context → Model → Tool → Observation → Next step
Practical example & important nuance

A research assistant gathers approved material, records its findings and stops when the brief meets its acceptance criteria.

A harness is more than “the loop.” Define stopping conditions, error handling and measurable completion.

Source [1]

02Foundation

Memory & state

Remembered information and current progress.

Memory retains selected information for reuse. State records the current task, completed actions, pending steps and execution status. A context window is temporary working context, not automatic durable memory.

Working context + Saved memory + Task state
Practical example & important nuance

Remember a permitted writing preference; separately record that a draft is awaiting review and has not been sent.

Cross-run memory must be implemented deliberately. Apply retention, deletion, freshness and user-isolation rules.

Source [2]

03Knowledge

RAG

Retrieve evidence before generating an answer.

Retrieval-augmented generation selects relevant external material and supplies it as context. Systems may use keyword, vector or hybrid search, with filtering and reranking.

Question → Retrieve → Select → Generate → Verify
Practical example & important nuance

Answer a property question from approved specifications; flag missing floor-area information instead of inventing it.

Retrieval does not guarantee truth or valid citations. Check relevance, source freshness and whether the cited passage supports the answer.

Source [2]

04Connection

MCP

A common interface to tools and context.

Model Context Protocol uses hosts, clients and servers to expose tools, resources and prompts. A compatible server and correctly scoped access are still needed.

Host → MCP client ⇄ MCP server → Tools / resources
Practical example & important nuance

An assistant retrieves an authorized document through a server without needing a custom model-specific integration.

“One standard plug” is a useful analogy, not universal compatibility. Governance by a foundation does not certify an individual server as safe.

Source [3]

05Knowledge

Skills

Reusable instructions and supporting resources.

In the Agent Skills format, a folder contains SKILL.md metadata and instructions, with optional scripts and reference files. Compatible hosts can load details when the skill is relevant.

Discover metadata → Load instructions → Use resources
Practical example & important nuance

A brand-review skill supplies tone rules and a checklist; the host supplies the tools and enforces access.

Progressive loading is a supported pattern, not a universal rule for every product. Skill scripts still need review and appropriate permissions.

Source [4]

06Assurance

Guardrails

Layered controls around decisions and actions.

Combine authorization, constrained tools, validation, sandboxing, review and monitoring. Prompt instructions alone cannot reliably enforce an action boundary.

Proposed action → Policy decision → Execute / block / review
Practical example & important nuance

Reading a record may be allowed automatically; deleting it can require a scoped permission and informed approval.

Least privilege and useful audit records are design requirements, not automatic properties. Logging must protect sensitive data.

Source [3]

07Assurance

Evals

Measure how well the whole system performs.

Use representative test tasks and explicit grading rules to assess outcomes, tool behavior and failure modes. Repeat trials and inspect traces to understand variability.

Task set → Run → Grade → Diagnose → Improve
Practical example & important nuance

Check whether an assistant found the correct record, avoided unauthorized changes and produced a supported answer.

Do not score only the final text. Combine reliable checks, calibrated model graders and human review; production monitoring complements prelaunch tests.

Source [5]

08Connection

A2A

A protocol for collaboration between agents.

Agent2Agent defines interactions using messages, tasks and artifacts. Agent Cards describe capabilities and interfaces so clients can discover how to interact.

Discover Agent Card → Request → Status / message / artifact
Practical example & important nuance

A coordinating agent requests a report from a separately operated research agent and tracks the task through completion.

An Agent Card is a description, not proof of competence or trust. Authentication and authorization still matter. A2A is not required for every multi-agent system.

Source [6]

09Connection

Multi-agent systems

Several agents coordinate toward a goal.

Agents may collaborate through orchestration, handoffs or peer interactions. Work can run in parallel, sequentially or in a mixed pattern depending on dependencies.

Assign → Collaborate → Resolve conflicts → Verify result
Practical example & important nuance

A researcher supplies evidence before a writer drafts; an independent reviewer then checks the claims.

Parallel specialists are one pattern. Extra agents add coordination cost and can repeat or amplify mistakes; compare with a single-agent baseline.

Source [1]

10Foundation · Added

Tool contracts

Define exactly what an action means.

A tool contract specifies inputs, outputs, permissions, side effects and error behavior. Structured arguments should be checked before a tool executes.

Schema → Validate → Authorize → Execute → Confirm
Practical example & important nuance

A CRM update requires a unique lead ID, allowed fields and confirmation that exactly the intended record changed.

A schema-valid call can still target the wrong record. Verify business rules, identity and the resulting state.

ImageFirm implementation guidance.

11Assurance · Added

Observability

See what happened across an entire run.

Correlate model requests, tool calls, timing, usage, errors and outcomes. Use traces to investigate failures and metrics to track trends.

Events → Correlated trace → Diagnosis → Eval case
Practical example & important nuance

A slow report can be traced to retrieval retries rather than blamed on the final generation step.

A trace is evidence about execution, not proof the output is correct. Redact secrets and restrict access to sensitive records.

Source [5]

12Assurance · Added

Budgets & recovery

Bound the run and recover without duplication.

Set time, cost and step limits; checkpoint progress where appropriate. Distinguish retryable failures from cases that need a person or a fresh plan.

Limit → Execute → Checkpoint → Retry / recover / stop
Practical example & important nuance

After a send timeout, check delivery status before sending again. Escalate if the result remains uncertain.

Repeated side effects are a major risk. Use idempotency where supported and verify previous completion before retrying.

ImageFirm implementation guidance.

The most common confusion

MCP, A2A and Skills do different jobs.

ConceptMain purposeTypical contentsDoes not establish
MCPConnect an AI application to capabilities and contextTools, resources, promptsAutomatic access to every service or trusted data
A2ASupport communication and delegated work between agentsAgent Cards, messages, tasks, artifactsVerified competence or permission to act
Agent SkillsPackage reusable task instructions and resourcesSKILL.md, optional scripts and reference filesA live service connection or an authorization grant

These mechanisms can coexist. “Skills” advertised in an A2A Agent Card describe capabilities; they are not necessarily Agent Skills folders containing SKILL.md. See sources [3], [4] and [6].

Apply the vocabulary

What are you trying to do?

Select a need to see which concepts are most relevant. This is a learning aid, not an automated architecture assessment.

RAG + context management

Retrieve relevant documents, preserve permissions and verify claims against the cited passages. A large context window alone does not ensure the right evidence is selected.

Add: source freshness checks, retrieval evals and an explicit “insufficient evidence” response.

Editorial audit

What the blueprint needed.

More precise definitions

The harness extends beyond a loop. Memory and state serve different purposes. Skills packaging depends on host support. MCP exposes resources, not an unrestricted data pipe.

Fewer implied guarantees

RAG can retrieve the wrong evidence. Evals must examine the whole run. Guardrails and logging require implementation. An Agent Card does not establish trust.

A more accurate collaboration model

Multi-agent work may be sequential or parallel and need not use A2A. MCP and A2A are complementary protocols, while Skills package reusable know-how.

Operational completeness

Tool contracts, observability, and budgets with recovery explain how agent behavior becomes measurable and manageable in real workflows.

Governance note: the Linux Foundation connection in the image is supported. MCP and A2A appear in the Agentic AI Foundation ecosystem; this does not certify products built on them. See source [7].

Questions to ask before launch.

Use this as a discussion checklist. Checked items indicate review, not certification. Selections are kept only for this page session.

0 of 10 reviewed

Sources & further reading.

Primary references reviewed for this edition. Protocol and product behavior may evolve. Definitions draw on the references below; examples and launch questions are ImageFirm editorial guidance.

  1. Anthropic · Building effective agents
    Agent workflows, orchestration and system design.
  2. Anthropic · Effective context engineering
    Working context, retrieval and persistent state.
  3. MCP · Architecture overview
    Hosts, clients, servers and protocol primitives.
  4. Agent Skills · Specification
    SKILL.md structure and progressive disclosure.
  5. Anthropic · Demystifying evals for AI agents
    Task outcomes, graders, trials and traces.
  6. A2A · Protocol specification
    Agent Cards, tasks, messages and artifacts.
  7. Agentic AI Foundation · Projects
    Project governance and ecosystem context.