Use EAP in a Workflow Pipeline ¶

AI-enabled work can use two complementary coordination models:

One Agent uses working knowledge, instructions, and tools to complete the whole job.
A workflow engine controls a fixed pipeline and assigns the most appropriate worker to each station.

Neither model is inherently better. Choose the model that fits the work.

When a deterministic pipeline needs an Agent at one of its stations, the workflow engine manages the pipeline and EAP provides the governed server-side Codex worker.

Choose The Coordination Model ¶

One Agent Completes The Whole Job ¶

Use one Agent when the work benefits from adaptive decisions across an end-to-end task. The Agent:

uses working knowledge, instructions, and tools
decides what to do next from the current evidence
adapts its approach as the work develops
produces the complete result

In the EAP operating model:

Codex is the digital team member doing the work.
AGENTS.md and supporting documents provide working knowledge.
An Agent Skill provides the instructions for a particular kind of work.
Scripts, CLIs, and MCP servers provide tools.

This resembles one person following assembly instructions to build a complete piece of furniture.

A Pipeline Coordinates Multiple Stations ¶

Use a pipeline when the route must be fixed and explicitly controlled. The workflow engine:

defines the pipeline and its stations
decides when work advances to the next station
gives each station one bounded task
selects the worker appropriate to that task

This resembles a production line that moves work through a defined sequence of stations.

flowchart TB
    subgraph WholeJob[One Agent completes the whole job]
        Request1[Work request] --> Agent1[Agent] --> Result1[Complete result]
        Knowledge[Working knowledge] --> Agent1
        Instructions[Instructions] --> Agent1
        Tools1[Tools] --> Agent1
    end

    subgraph Pipeline[A pipeline coordinates multiple stations]
        Request2[Work request] --> Engine[Workflow engine]
        Engine --> Station1[Station 1] --> Station2[Station 2] --> Station3[Station 3]
        Station3 --> Result2[Complete result]
    end

Both models are valid. Choose the coordination model that fits the work.

Use The Simplest Reliable Worker At Each Station ¶

Not every station needs an Agent. Select the simplest worker that can perform the bounded task reliably.

Code ¶

Use code for exact, repeatable computation, validation, or system action, including:

data transformation and calculation
format and policy validation
API calls with explicit inputs and outputs
system changes that must follow strict rules
post-condition checks, retries, rollback, and compensation actions

Direct LLM Call ¶

Use one direct call through Cisco Collab LLM Proxy when the input is already complete and the station needs one bounded inference, including:

sentiment analysis
text classification or routing
structured-data extraction from fixed input
summarization or rewriting of prepared content

These stations do not need an Agent Harness.

EAP Agent ¶

Use an EAP Agent when the station needs working knowledge, adaptive investigation, multiple tool calls, or decisions based on intermediate results, including:

collecting and correlating evidence across systems
investigating a likely root cause from logs, metrics, recent changes, and runbooks
deciding which additional evidence to query when information is incomplete
using team knowledge to produce an evidence-backed recommendation

The workflow engine invokes the server-side Codex Agent through the EAP API.

Human ¶

Use a human for:

decisions for which a person must retain final accountability
exceptional cases that models and rules cannot handle reliably
gates that policy explicitly reserves for a person

flowchart LR
    Station[Pipeline station] --> Code[Code
Exact and repeatable]
    Station --> LLM[Direct LLM call
One bounded inference]
    Station --> Agent[EAP Agent
Adaptive investigation and tool use]
    Station --> Human[Human
Accountability and exceptions]

A pipeline does not need an Agent at every station.

Divide Responsibility Between The Workflow Engine And EAP ¶

The workflow engine and EAP govern different execution boundaries.

The Workflow Engine Runs The Pipeline ¶

The workflow engine is the top-level orchestrator. It owns:

station definitions and execution order
events, schedules, and API requests that start the workflow
pipeline state and station transitions
decisions to continue, retry, pause, recover, or stop
workflow-level checks and human-in-the-loop gates
the end-to-end execution record for the complete workflow

The workflow engine answers:

Where is this pipeline now, and which station should run next?

EAP Runs The Agent At A Station ¶

When a station needs an Agent Harness, the workflow engine calls a server-side Codex Agent through the EAP API. EAP owns the Agent-execution boundary:

running Codex in a governed server-side environment
using the Agent's non-human identity instead of an engineer's identity
providing the working knowledge, Agent Skills, and tools from the customer-owned agent repo
constraining the Agent to approved MCP servers, credentials, and network destinations
applying Agent tool-call human-in-the-loop rules
recording the caller, Agent, tool calls, approver, and Agent-run result
returning the station result to the workflow engine

EAP answers:

Who called this Agent station, what did the Agent do, which tools did it use, and what result did it return?

flowchart LR
    Trigger[Event / schedule / API request] --> Engine[Workflow engine
Pipeline state and transitions]
    Engine -->|Invoke Agent station| EAP[EAP
Governed Agent execution]
    EAP --> Codex[Server-side Codex]
    Codex --> Tools[Approved tools]
    Tools --> Codex
    Codex --> EAP
    EAP -->|Return station result| Engine
    Engine --> Next[Next station]

EAP does not own pipeline state or station transitions.

Example: A SaaS Incident-Response Pipeline ¶

Consider a production incident-response pipeline triggered by a monitoring alert or customer report. The workflow engine advances the complete pipeline, and each station uses the worker best suited to its bounded task.

flowchart LR
    Trigger[Alert or
customer report]
    S1[1. Normalize and enrich
Code]
    S2[2. Classify customer impact
Direct LLM]
    S3[3. Investigate likely root cause
EAP Agent]
    S4[4. Validate remediation plan
Code]
    S5[5. Approve production change
Human]
    S6[6. Execute and verify remediation
Code]
    S7[7. Draft incident update
Direct LLM]

    Trigger --> S1 --> S2 --> S3 --> S4 --> S5 --> S6 --> S7

Station 1: Normalize And Enrich — Code ¶

Code parses the alert, identifies the service, environment, region, and time window, and collects predefined telemetry and recent changes.

The fields, rules, and queries can be defined exactly, so this station does not need an LLM or Agent.

Station 2: Classify Customer Impact — Direct LLM ¶

Cisco Collab LLM Proxy uses customer comments or issue descriptions to classify sentiment, impact, or urgency.

The input is complete and the task needs one bounded inference, so this station does not need an Agent Harness.

Station 3: Investigate Likely Root Cause — EAP Agent ¶

The workflow engine invokes an incident-investigation Agent through the EAP API. This station uses an EAP Agent because Codex may need to:

read the team's AGENTS.md, runbooks, and service documents
analyze evidence collected by earlier stations
decide which additional logs or metrics to query
query other systems through approved MCP servers
compare possible causes and produce an evidence-backed recommendation

EAP manages this Agent run. The workflow engine waits for the station result and decides whether to advance the pipeline.

Station 4: Validate The Remediation Plan — Code ¶

Code checks whether the Agent's recommendation is an allowed remediation, its required parameters are complete, its target resources are within the incident scope, and its preconditions are satisfied.

The Agent supplies an investigation result and recommendation. Code decides whether that recommendation satisfies executable rules.

Station 5: Approve The Production Change — Human ¶

When policy requires a human decision, the workflow engine pauses the pipeline and waits for an authorized operator to approve or reject the production change.

This is a workflow-level human-in-the-loop gate.

Station 6: Execute And Verify Remediation — Code ¶

Code performs the validated and approved rollback, restart, configuration change, or other remediation. It then checks service health, error rate, latency, and other post-conditions.

Code makes idempotency, retry, partial-failure, and rollback behavior explicit.

Station 7: Draft The Incident Update — Direct LLM ¶

The workflow engine sends confirmed facts and results to Cisco Collab LLM Proxy to draft a customer-facing or internal status update.

The facts are complete and the task only generates text, so it normally does not need an Agent Harness.

Use an EAP Agent only where a station needs working knowledge, adaptive investigation, and tool use.