ImageFirm / INTELLIGENCE
FIELD GUIDE · OCT 2026
Strategy / Technology / Execution

Understand AI.
Build with judgment.

12 concepts for turning capable models into useful, measurable and accountable systems.

Rebuilt from your nine-concept blueprint. Corrected where the claims were too broad. Extended where production essentials were missing.

Explore the field guide ↓

The systems behind the intelligence.

12 concepts

NO.THE CONCEPTHOW IT WORKS / WHAT TO CHECK
01

Agentic loops

Execution
Give autonomy a clear finish line.

An agent selects actions, observes results and adjusts its next step. A workflow follows a predefined path; many useful systems combine both.

Goal → Context → Act → Check → Continue / stop
Implementation check & editorial note

Put it into practice. Set success criteria, time and spend limits, maximum steps, and a route to a person when progress stalls.

Example: research a property brief, verify claims, then stop when the evidence checklist is complete.

Source [1]

02

Model Context Protocol

Integration
A shared interface, not universal access.

MCP connects a host application through clients to servers exposing tools, resources and prompts. Compatible servers and explicit access are still required.

Host ⇄ MCP client ⇄ MCP server ⇄ Service
Implementation check & editorial note

Put it into practice. Review server trust, authorization scope and data access. MCP does not make a connected service safe by itself.

Correction: “one protocol, every tool” overstates compatibility; permissions require implementation and enforcement.

Source [2]

03

Subagents & multi-agent systems

Execution
Split work only when the split helps.

An orchestrator can delegate bounded work to specialist agents, then reconcile their outputs. Separate contexts can help focus, but coordination adds overhead.

Orchestrator → Specialists → Reconciliation
Implementation check & editorial note

Put it into practice. Define task boundaries, shared evidence, ownership and a conflict-resolution rule. Compare against a single-agent baseline.

Correction: more agents can increase cost and failure points; they do not guarantee higher quality.

Source [1]

04

AI gateways

Integration
Centralize traffic policy.

A gateway can route model requests and apply authentication, quotas, logging and fallbacks. Tool routing is an optional capability, not an inherent feature.

Applications → Gateway policy → Model providers
Implementation check & editorial note

Put it into practice. Check provider feature differences, timeout behavior and data handling. A fallback must preserve required capabilities.

Correction: a gateway is not an agent runtime, and open-weight models are not automatically open-source software.

ImageFirm implementation guidance.

05

Inference economics

Measurement
Measure cost per successful task.

Total cost includes model usage, tools, retries, infrastructure and human review. Cached input, output and reasoning charges depend on the model and provider.

All operating cost ÷ Accepted outcomes
Implementation check & editorial note

Put it into practice. Track quality and latency alongside spend; cheap calls can become expensive workflows. Include failed attempts in the numerator.

Correction: “cache hits cost 5%,” “output costs 5×,” and “batch is 50% off” are not universal rules. Check current rates.

Source [3]

06

Evaluations

Measurement
Test the outcome and the process.

Use representative tasks, explicit rubrics and repeat trials. Combine deterministic checks, calibrated model grading and human review as appropriate.

Test set → Run → Grade → Compare → Regressions
Implementation check & editorial note

Put it into practice. Measure task success, tool correctness, harmful actions, cost and latency. Keep a held-out set and inspect failures.

Correction: passing an evaluation does not mean “right every time.” Graders can be wrong and production inputs can differ.

Source [4]

07

Guardrails & permissions

Trust
Enforce boundaries outside the model.

Combine scoped access, validated tool inputs, sandboxing and approvals for consequential actions. Treat retrieved content and tool output as untrusted data.

Proposed action → Policy check → Authorize / block
Implementation check & editorial note

Put it into practice. Use least privilege and protect secrets. Check the actual action and its target before execution; output filtering alone is insufficient.

Correction: a green “safe output” box is an intended result, not a security guarantee.

Source [2]

08

Observability

Measurement
Make failures explainable.

Capture correlated traces, tool calls, timings, usage, errors and outcomes. Inspect where a run diverged from its objective.

Run → Trace + metrics + feedback → Diagnosis
Implementation check & editorial note

Put it into practice. Redact personal data and secrets; set retention and access rules. Feed recurring failures into the evaluation set.

Correction: logging everything can expose sensitive data. Trace useful events under a clear data policy.

Source [4]

09

Agent harness

Execution
The operating environment shapes performance.

The harness manages context, tools, execution state, permissions, checks and recovery around the model. Its design can materially change results.

Model + Context + Tools + State + Controls
Implementation check & editorial note

Put it into practice. Version prompts, tools and model configuration together; test the full system after changes.

Correction: the blueprint’s benchmark ranking claim lacks a verifiable setup and citation; it has been removed.

Source [1]

10

Context, retrieval & memory

Integration · Essential addition
Supply relevant evidence at the right time.

Retrieval-augmented generation (RAG) fetches external material into context. Memory stores selected state across steps or sessions; neither changes model weights by itself.

Retrieve → Check source → Add context → Answer
Implementation check & editorial note

Put it into practice. Track provenance and freshness, respect document permissions, and allow outdated memory to be corrected or deleted.

Essential addition: retrieval can fail or surface conflicting evidence. Cite sources and acknowledge missing information.

Source [5]

11

Reliable tool execution

Trust · Essential addition
Make actions recoverable.

Validate structured arguments, set timeouts and distinguish transient errors from permanent failures. Record completed actions before deciding whether to retry.

Validate → Execute → Confirm → Record / recover
Implementation check & editorial note

Put it into practice. Use idempotency keys where supported to prevent duplicate side effects. Never blindly retry payments, messages or record creation.

Essential addition: a successful HTTP response is not always a successful business outcome. Verify the resulting state.

ImageFirm implementation guidance.

12

Human oversight & ownership

Trust · Essential addition
Name the person responsible.

Assign an owner for the workflow, its data and its acceptance criteria. Define when automation must stop and what a reviewer needs to see.

Owner → Risk limits → Review → Escalation
Implementation check & editorial note

Put it into practice. Give reviewers the proposed action, evidence, uncertainty and consequences. Maintain a practical rollback or remediation path.

Essential addition: human approval is only useful when it is informed, timely and recorded.

ImageFirm implementation guidance.

Blueprint audit

What changed—and why.

Remove false precision

Fixed caching discounts, output/input price ratios and blanket batch savings were replaced with provider-dependent pricing. The uncited harness benchmark ranking was removed.

Separate mechanisms from guarantees

MCP needs compatible implementations. More agents do not assure better work. Evals estimate performance; guardrails reduce risk. Neither proves universal reliability or safety.

Make the stopping rules explicit

A run stops on verified completion, budget or time limits, repeated failure, or escalation. “Done” must have observable acceptance criteria.

Add the operational foundations

Context and retrieval, recoverable tool execution, and accountable human oversight complete the original nine-concept map.

Interactive example

What does a successful outcome actually cost?

Use the same period and currency for all inputs. Count accepted outcomes against a defined quality standard.

Illustrative defaults, not market benchmarks. Operating cost includes failed attempts and review. Setup costs and taxes are excluded unless you add them.

$3.75
$300 total ÷ 80 accepted outcomes
From understanding to delivery

Your minimum launch checklist.

These checks support a review; completion is not a certification. Selections last for this page session only.

0 of 10 reviewed

Sources & reading.

Primary documentation checked for this edition. Pricing and protocol details can change; follow the source for current terms. Sources support the technical explanations; the checklist and examples are editorial recommendations.

  1. Anthropic · Building effective agents
    Workflows, agents and orchestration trade-offs.
  2. MCP · Architecture overview / Security considerations
    Clients, servers, protocol primitives and authorization.
  3. OpenAI · API pricing / Prompt caching
    Model-specific rates and cache behavior.
  4. Anthropic · Demystifying evals for AI agents
    Outcome grading, traces and evaluation design.
  5. Anthropic · Effective context engineering
    Context selection, retrieval and state management.