Resource
The Agent Operating Model
Everybody sells "an AI agent" today. Very few can tell you how it decides, what it reads, and who signs off before it touches your money. This is the operating model we hold AI deployments to: prompt engineering, context engineering, semantic networks, and harness engineering, traced through one warranty claim from the first phone call to the committed filing.
Agent Operating Model
Prompt engineering|Context engineering|Semantic networks|Harness engineering
Following one warranty claim end to end: a customer reports a failing part, the agent drafts the claim, pulls the warranty terms and install records — traversing the knowledge graph that links them — then gates and commits the filing and replacement order.
Prompt
engineering
engineering
The message
Compose the instruction, constraints, examples, and required output.
Role
“warranty specialist”
Task / instructions
“file the claim”
Constraints
in-warranty only
Examples
a filed claim
Output schema
claim fields
Acceptance criteria
serial attached
1
Compose
PROMPT
"You are a warranty specialist…"
"File a claim for this part…"
Example: prior claim → filed
Respond in these fields…
Done when serial + proof attached
"File a claim for this part…"
Example: prior claim → filed
Respond in these fields…
Done when serial + proof attached
one assembled input
2
Send
LLM
inference
inference
3
Generate
Response
raw output
4
Refine — adjust the weakest ingredient
Unit of work: one input.
Models & prompt tools
AnthropicClaude
OpenAIChatGPT
MicrosoftMicrosoft Copilot
GoogleGemini
AmazonAmazon Q / Amazon Nova
IBMwatsonx.ai Prompt Lab
SalesforceEinstein / Agentforce
Mistral AIMistral (Le Chat)
DatabricksAI Playground
OracleOCI Generative AI
AmazonAmazon Q / Amazon NovaContext
engineering
engineering
The evidence
Select authoritative, current, minimal evidence and state for each step.
User query“part failed early”
System promptwarranty advisor
Retrieved docswarranty terms
Tool outputsinstall date: 2023
Memorycustomer: Rivera
Prior turnsclaim started
1
Gather
Curator
select, rank, compress, drop
2
Curate
Context window
chunk
chunk
…
drop
fits the budget
3
Feed
LLM
inference
inference
4
Step out
Response or
tool call
tool call
Source authority
Authoritativesystem of record, approved policy
Workingcurrent task notes, recent tool output
Recallhistory & agent memory
Untrustedweb, attachments, model-generated
Context engineering decides not only what fits, but what is authoritative, current, and safe to use.
5
Gated write — persist only durable, source-backed, approved facts; expire the rest
Unit of work: what stays in the window each step.
Memory & retrieval
AnthropicMCP (protocol)
OpenAIResponses API: Conversations + File Search
MicrosoftAzure AI Search
GoogleVertex AI Agent Search
AmazonBedrock Knowledge Bases
IBMWatson Discovery
SalesforceData Cloud RAG
Mistral AIMistral Libraries
DatabricksDatabricks AI Search
OracleGen AI Agents RAG tool
AmazonBedrock Knowledge BasesSemantic
networks
networks
The knowledge graph
Model the domain as typed entities and relationships the agent can traverse — not loose text.
EntitiesRivera, compressor
Relationshipsowns, covered-by
Ontology / schemawarranty domain
Records & docsinstall log, terms
Embeddingssemantic index
1
Extract
Graph builder
entity & relation extraction
2
Build
Knowledge graph
a traversable domain model
3
Query
Retrieved
subgraph
subgraph
Compressor ▸ Warranty ▸ in‑warranty
grounds context & the answer
Graph elements
Entitya thing — customer, part, policy
Relationshiptyped edge — owns, covered-by
Attributeinstall date, coverage window
The graph answers questions loose documents can't: is this exact part, on this unit, still inside its coverage window?
4
Enrich — write new verified facts back as nodes and edges; keep the graph current
Unit of work: an entity and its relationships.
Graph & semantic layer
Neo4jNeo4j + GraphRAG
AmazonAmazon Neptune
MicrosoftGraphRAG
GoogleSpanner Graph
OracleOracle Graph
IBMwatsonx.data
DatabricksUnity Catalog
OntotextGraphDB
StardogStardog
TigerGraphTigerGraph
AmazonAmazon NeptuneHarness
engineering
engineering
The operating system
Gather, act, verify, approve when needed, commit, observe, and retry safely.
User request“part failed early”
Memory LLM/DBcustomer: Rivera
Prior tool outputsinstall date: 2023
Retrieved docswarranty terms
HARNESS — full-stack cognitive architecture
1Gather
Curator
Window
fits budget
context engineering zone
2Act
PROMPT
"You are a warranty specialist…"
"File the claim…"
serial + terms…
Respond in fields
"File the claim…"
serial + terms…
Respond in fields
send
LLM
Tools
exec, fetch
Sub-agents
specialists
prompt engineering zone · one assembled input
3Verify & commit
Verifier
schema · tests
API readback
LLM judge
API readback
LLM judge
pass
4Approve gate
Green — auto
Yellow — owner approval
Red — Owner approval
Commit / send
5
Retry — re-run gather → act → verify with updated context
Verifier stack
Deterministic checks · External readback · LLM review · Human approval
Action gates
Green — read, draft, classify, retry
Yellow — owner approval before write
Red — Owner approval to send, buy, use credentials, or write to production
Retries may self-correct analysis and preparation. Commit actions require the right authority.
Unit of work: a bounded job — intent, evidence, tools, verification, and a stop condition.
Agent frameworks
AnthropicClaude Agent SDK
OpenAIOpenAI Agents SDK
MicrosoftMicrosoft Agent Framework
GoogleADK + Vertex AI Agent Engine
AmazonBedrock Agents
IBMwatsonx Orchestrate
SalesforceAgentforce
Mistral AIMistral Agents API
DatabricksDatabricks Agent Framework
OracleOCI Gen AI Agents + ADK
AmazonBedrock AgentsSwipe sideways to explore the full diagram, or download the PDF above.
This is the standard we hold AI to.
If a tool can't explain what it read, why it decided, and who approved the action, it doesn't belong near your customers or your bank account. We help trades businesses pick agents that pass this bar and coach them until they earn more autonomy.
Talk to an advisor