liveBook · Manning

Agentic workflow

As explained in AI Agents for Offensive Security

A workflow in which AI agents and security tools pass information and decisions through ordered stages.

An agentic workflow is a controlled pipeline in which AI agents, security tools, people, and safety controls move work through ordered stages. Rather than letting one broadly capable agent pursue an opaque task, the workflow assigns bounded decisions to specialized stages and preserves the information behind each transition. section 1.0 section 3.0

Artifacts are the durable interface of the workflow. Port scans, HTTP probes, enumeration results, behavioral observations, model responses, and run information become structured outputs that later stages can consume. Treating these outputs as first-class objects lets reconnaissance inform triage, triage inform prioritization, and prioritization direct testing, reporting, or escalation. section 1.0 section 3.1

A defensible agentic workflow separates these responsibilities:

  • Specialized agents reason within a defined role, act through tools, and record their behavior as artifacts. section 3.0
  • Orchestration controls execution order, dependencies, and failure handling instead of embedding those concerns in agent reasoning. section 3.0
  • Safety gates and human approval constrain escalation and keep objectives and scope under explicit control. section 1.0 section 3.0
  • Logging, tracing, and recorded transitions make decisions observable, reproducible, and easier to diagnose when a stage fails. section 3.0 section 3.1

The pattern contrasts with both conventional linear automation and a single “smart” agent. Scripts and frameworks accelerate execution, while agents can interpret results, prioritize effort, and assist decisions. A single agent may collapse dependent work into brittle behavior that is difficult to inspect or defend. Staged pipelines instead use explicit artifact contracts and control points so that each step does less, but the system is clearer overall. section 1.0 section 3.0

Reconnaissance-to-escalation pipeline

A security engagement can route reconnaissance observations to a triage agent, pass filtered findings to prioritization, and then direct the selected work to testing, reporting, or escalation. Each handoff records artifacts; orchestration determines when the next stage may run; approval and safety checks intervene before riskier actions. The result is a cohesive workflow whose later evidence can be traced to earlier decisions. section 3.1 section 3.0

Figure 3.1 The general multi-agent system. Each agent consists of its own reasoning loop and can operate either independently or be coupled to other modules. Each agent can generate artifacts, which contain the model responses and information about the model run. These artifacts connect the agents, such that the outputs of one agent become the inputs of another. This structured pipeline is designed to achieve a specific goal defined by the user. To ensure our AI systems don’t go off the rails, guardrails such as safety gates and logging/tracing are implemented for observability.

The diagram makes the workflow boundary visible: independent or coupled reasoning loops exchange artifacts, while safety gates and logging/tracing surround the pipeline for observability. section 3.1

Agentic workflows use AI as decision support, not unsupervised execution. Tools perform actions; humans set objectives and approve escalation; the pipeline governs information flow. This division supports authorized security testing while retaining accountability for what was decided, executed, and observed. section 1.0 section 3.0

How AI Agents for Offensive Security covers it

How deep it goes
Explains staged composition, artifact interfaces, orchestration, failure handling, safety gates, and observability rather than agent internals.
Examples
Uses reconnaissance observations flowing through triage and prioritization into testing, reporting, or escalation.
What it assumes
Assumes security practitioners already know tools while introducing AI agents and pipeline-based information flow.
What you can do after
Build it from scratch as a multi-agent pipeline with explicit artifact handoffs, controls, and human oversight.
How it connects
Connects AI agents with artifacts, orchestration, safety gates, human approval, logging, tracing, and error handling.

Also explained in

Compare how 4 books explain agentic workflow →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Agents for Offensive Security. The section links point to the parts of the book it draws on.