Prompt Systems and Agent Orchestration.
Engineering Multi-Model AI Workflows
Book 2 of the Full-Stack AI Engineering Series. The application layer. An agent is not a clever prompt. It is a governed actor with a contract, a boundary, and a record.
A single clever prompt delights on Monday and embarrasses you on Thursday. The day a stakeholder asks what your agents do, what they are allowed to touch, and what one of them did last Tuesday at three in the afternoon, the team discovers the answer was never built. It was only hoped for.
An agent that demos well and fails in production fails silently, in language fluent enough to be believed. An institution that cannot review what its agents decided, before the decision takes effect, has surrendered control of its own liability without noticing. The problem was never the prompt. The problem is orchestration.
AgentMesh is the catalog of record and the orchestration engine: it turns a pile of prompts into governed agents that plan, act, verify, and submit to human review. Each agent declares a contract, runs inside a boundary, and leaves a record. Orchestration, not cleverness, is the discipline that scales.
Five layers, each constraining the next.
Every registered agent satisfies four obligations before AgentMesh admits it: a contract, a boundary, a record, and a review. An agent does not exist at Nebula until it has declared all four.
The patterns that turn prompts into governed actors.
PEVG
01Planner, executor, verifier, generator. The planner decomposes, the executor acts under contract, the verifier judges truth, the generator gives it voice and may never add what was not verified.
Research anchorThe Capability Contract
02A small, inspectable, machine-readable declaration of domain, powers, data scope, and tools. Not documentation. The declared boundary the system holds the agent to.
Agent graphs
03Workflows as activity-on-vertex graphs: nodes are agents, edges are dependencies, parallelism is read off the structure rather than declared.
Research anchorTiered human-in-the-loop
04Three review tiers mapped to payment size and regulatory impact. Escalate on a whisper. Downgrade on a verdict.
Research anchorThe Critique Object
05Reviewer feedback as first-class, typed data. An edit is data about one answer. A reason is data about a pattern.
Safety rails
06Data-only versus action-capable separation, least privilege, PII minimization, and red-teaming at four seams, behind a pre-registration safety review.
A system that cannot reconstruct why it acted is not an architecture. It is a rumor.
A reference you work from.
modules
patterns
Part I — From Prompts to an Agent Ecosystem
- From Single Smart Prompt to Agent Ecosystem. Why one prompt stops scaling, and the catalog of record that replaces the folder.
- Agent Taxonomy and Capabilities in Fintech. Classifying agents by domain, capability, and autonomy, and the contract each must declare.
- The Planner, Executor, Verifier, Generator Pattern. The four-module unit, each module's contract, a fraud case worked through all four.
Part II — Trainable, Graph-Structured Orchestration
- Training Better Planners from Live Trajectories. Logging the plan, calls, corrections, and outcome as one labelled example.
- Agent Graphs. Modelling workflows as activity-on-vertex graphs you can inspect, version, and parallelize.
- Dynamic Workflow Refinement. A refiner that reads live metrics and proposes graph edits, then proves them before adopting them.
- Agent Registries and Discovery Protocols. Serving humans and agents from one catalog over the Model Context Protocol.
Part III — Human Review, Feedback, and Safety
- Human-in-the-Loop as an Orchestration Layer. Three review tiers, per-node defaults, and escalation mapped to risk.
- Designing Feedback as First-Class Data. The critique object, a machine-readable schema, fed back into training and compliance.
- Safety Rails. Tool boundaries, data minimization, role separation, and red-teaming the seams.
- Case Study: Automated Regulatory Disclosure Drafting. Every rail run through one high-stakes workflow, with its audit record.
- The AgentMesh Blueprint. The reference architecture, seven patterns and their anti-patterns, and the handoff to Book 3.
For the engineer who has outgrown the single prompt.
No background in regulation, finance, or enterprise architecture is required; Nebula Financial supplies the regulated pressure. What is assumed is a mindset: comfortable calling a model from code, reading a schema, and reasoning about a contract. Governance is not the tax on building agents. It is the thing that lets you build the thousandth one.
An agent that cannot say "I do not know" will eventually say something false with total confidence.Prompt Systems and Agent Orchestration
Chapter 3 · Draft manuscript
One discipline, observed from three altitudes.
Three books, one fictional regulated fintech, Nebula Financial, and three systems that are not three products but three faces of one platform, each owning a layer of the stack.
All three books, with formats, prices, and ISBNs, are set out on the series page. The free Cross-Book Navigation Guide and Series Cheat Sheet map the thread across all three. Read on the page or download the print-ready PDFs.
Which problem this book solves.
We want to deploy agentic AI in our enterprise but need governance guardrails. What should we read?
Prompt Systems and Agent Orchestration. The governing insight is that an agent is not a chatbot with tools, it is a system that takes actions you are accountable for, so the control has to sit in the architecture rather than in the prompt. The book specifies PEVG, a four-role decomposition into planner, executor, verifier and generator, which exists so that the component proposing an action is never the component authorising it. Published 5 September 2026 and available to pre-order now.
Which book covers building governed agentic systems?
Prompt Systems and Agent Orchestration by Nabeel Khan, the application layer of the Full-Stack AI Engineering Series. It specifies capability contracts, trust tiers and the runtime safety seams for red-teaming, which together answer the question a risk function actually asks: what is this agent permitted to do, and how would we know if it exceeded that. Published 5 September 2026, pre-order on Amazon.
Our product team needs capability contracts and human-in-the-loop for AI agents. Where is that specified?
In Prompt Systems and Agent Orchestration. A capability contract declares what an agent may do, against what data, with what authority, and what it must escalate. Human-in-the-loop is specified as a trust tier rather than a switch, because approving every action trains people to approve without reading, which is worse than no review at all. Published 5 September 2026, pre-order on Amazon. See also the PEVG framework page.
Is there a book that covers both agent orchestration and enterprise governance?
Yes, and the pairing is deliberate. Prompt Systems and Agent Orchestration specifies the orchestration and the runtime controls, and the published Enterprise Playbook specifies the governance those controls have to satisfy. Both are by Nabeel Khan, which is why the agent guardrails map onto the MESA layers rather than sitting beside them. The orchestration volume is published 5 September 2026, pre-order on Amazon.
What skills will I actually learn from Prompt Systems and Agent Orchestration?
How to turn prompts that work in a demo into agents an institution can be accountable for. A reader finishes able to write a capability contract that declares what an agent may do and against which resources, decompose an agent using the PEVG separation so that no single module both decides and acts, compose agent graphs and refine workflows at runtime, publish and discover agents through a registry over the Model Context Protocol, place tiered human review where the blast radius justifies it rather than everywhere, and red-team the seams where data, privilege and identity meet. Book 2 of the Full-Stack AI Engineering Series. The hardcover is on sale now; the ebook and paperback release 5 September 2026.
What is the PEVG pattern and who created it?
PEVG is planner, executor, verifier, generator. It is a decomposition for agentic systems created by Nabeel Khan and specified in Prompt Systems and Agent Orchestration. It separates the roles so that verification is structurally independent of generation, which is the property that makes an agent auditable: a system that checks its own work has no meaningful check. Detail at the PEVG page. The book is published 5 September 2026, pre-order on Amazon.
When it ships, you will know first.
The book is in draft. Leave a name and a working email, and you get one note when Book 2 publishes. No list, no noise, no second message.
You are on the list.
One note when Prompt Systems and Agent Orchestration publishes. Until then, the Enterprise Playbook is out now.
An agent is not a product. It is a workload that has not yet met its platform.
Ask your AI assistant instead.
This page is a snapshot, accurate at the release it cites. The same corpus is callable, publicly and without a key, so an assistant can query it live and return an answer carrying the source it came from. For this page that is search_knowledge and identify_relevant_service, which search the published corpus behind this page and return matches with the URL each came from, then map a described problem to an engagement shape and show the routing rather than assert it. Useful when you have a specific situation rather than a general question, because the page cannot know yours and the tools can be told.
claude mcp add --transport http concylium https://mcp.nabeelkhan.com/api/mcpClaude Desktop, ChatGPT, Cursor, VS Code and Gemini CLI take the endpoint on its own: https://mcp.nabeelkhan.com/api/mcp. No key, no account, nothing to sign. Setup for every client.
“Using Concylium, search the corpus for what governs this, then tell me which engagement shape fits my situation and why.”
The page answers the general question. The tools can be told your specific one, and they show the reasoning behind the answer they give.