PROCESSES · AGENT STEPS · HUMAN APPROVALS

One process.
Agents work.
People approve.

Build a process from individual steps. An agent receives a clear assignment, completes its task and hands over the result. Wherever approval is required, the assigned person must review and decide before the process continues.

Isometric illustration: watcher, agent, script, test and human work one after another; a dashed path leads from approval back to rework.
Illustration · dashed: sent back for rework, then reviewed again.

A request becomes a shared process.

The process connects several tasks to one goal. Agents prepare the work; human approval is a separate step in the workflow.

  1. 01Request
  2. 02Agent analyzes
  3. 03Agent drafts
  4. 04Person approves
  5. 05Next step

Understand a request, identify missing details and prepare a draft reply.

Awaiting approval · example state

Draft reply is ready

  • Agent step 1: analyze the request.
  • Agent step 2: draft a reply.
  • Approval step: Lea reviews and decides.

Every agent step has a clear assignment.

One agent can handle several steps; several agents can work in sequence. Define the inputs, task, expected result and boundaries for each step.

  1. 01Request
  2. 02Agent analyzes
  3. 03Agent drafts
  4. 04Person approves
  5. 05Next step
Agent
Fink handles the request using a suitable AI model.
Step 1 · Analysis
Summarize the request and identify open questions.
Step 2 · Draft
Use the analysis to prepare a proposed reply.
Handover
Send the result to the assigned person for review.
A completed agent step is not yet a human approval.

A person must decide at the approval step.

In this example, the process waits here. Lea sees the assignment and the results of the agent steps. Only her explicit approval allows the next defined step.

  1. 01Request
  2. 02Agent analyzes
  3. 03Agent drafts
  4. 04Person approves
  5. 05Next step
Lea · assigned approver

This approval step is assigned to Lea.

Awaiting human approval

Is the reply accurate and complete?

ApproveAllow the next defined step
Return for reworkSend feedback to the agent
RejectStop or follow the defined rejection path
Explore roles and approvals →

Rework is part of the process.

When details are missing or a draft needs changes, the task goes back with specific feedback. The agent revises it; the new result must be reviewed and approved again.

  1. 01Request
  2. 02Agent analyzes
  3. 03Agent drafts
  4. 04Person approves
  5. 05Next step
Approval

Continue with the approved result

Rework

Agent revises → person reviews again

Error / rejection

Pause, resolve or follow the defined path

Tasks, results and decisions belong together.

A process run shows the current step, the responsible agent and where a human decision is needed. The connection between assignment and result stays clear.

  1. 01Request
  2. 02Agent analyzes
  3. 03Agent drafts
  4. 04Person approves
  5. 05Next step
Reply to a customer request
Analyze → draft
Lea reviews the draft reply
Hand over the approved result
Assignment, results, feedback and decision
Explore process runs and pending approvals →

Development status: These examples explain the current product structure and do not run a process. The initial scope focuses on processes, agent steps and human approvals. The new process engine with enforced approval pauses is being built; supported steps are checked for each test version.

The whole process counts. People included.

Teams and public administrations plan processes around tasks, responsibilities and a clear result. A workflow built entirely around AI can overlook human questions and decisions. Managing human tasks alone does not automatically connect technical steps and their costs either. A shared workflow brings work, handoffs, review, approval and rework together.

Tools such as n8n and Make connect applications and can include human approvals. A complete working process still requires deliberate design of human tasks, responsibilities, feedback and handoffs. c9n starts with building that shared process: What does an agent handle, what does a person work on, and who takes responsibility for the result?

What does a task really cost? Human time spent on work, review and rework × internal hourly rate + AI, infrastructure and allocated tool costs. Total turnaround time, including waiting, and result quality are considered separately. This provides a way to compare a manual workflow with a shared human-and-AI workflow for the same task.

Illustration with example values: €60 internal hourly rate, €1.50 shared AI, infrastructure and tool costs. You can enter your own numbers below.

Calculate with your numbers

The values are examples. Replace them with your own; the result updates immediately. Waiting time and quality are not included.

Before · minutes per task
With AI · minutes per task

Before 40 min · €40.00

With AI 25 min · €26.50

Saving per task: 15 min · €13.50. At 100 tasks per month: 25 h · €1,350.00.

A measurement approach for process planning. The website uses fictional calculations; this does not promise automatic tracking of these costs.

Explore the local structure ↓

Which process would you like to build together?Discuss your process →
How to build a process ↓

PROCESSES · AGENT STEPS · HUMAN APPROVALS

Build the workflow. Define the approvals.

Turn a goal into individual tasks: define the sequence, the right agent and the points where a person must decide.

01

Build the process

What is the goal? What starts the process, which steps are needed and what result should it produce?

Example: request → analysis → draft → approval.

02

Define agent steps

Assign each step to an agent with a task, inputs and an expected result. Scripts and tests can add predefined tasks.

Example: Fink analyzes the request, then prepares a draft.

03

Add human approval

The assigned person reviews the result. Approval allows the process to continue; returning it leads to rework and another review.

Example: Lea must approve the draft reply before the workflow continues.

LOCAL VERSION · ORIENTATION

How the local version is organized.

Build processes in Workflows and open the agents involved. The Agents area brings their jobs and controls together. The GPU pool organizes computing capacity for their models. Settings is where you configure connections to other services. This overview explains the intended structure of a local c9n installation.

01

Agents

From a shared process to an individual task: the process definition in Workflows links to the agents involved. People review results and give the required approvals.

Overview
See actors, their responsibilities and the current work status together.
Processes
Define a process in Workflows, review its actors and jobs, and open an agent to define its steps. Assign human approval steps to a responsible person within the process.
Jobs
Follow waiting, running and completed tasks. A job belongs to the agent carrying it out and to the relevant process.
Agent details and history
Understand an agent’s assignment, goals and limits, and review available results, feedback and recent runs.
Controls and schedules
Define when an agent should work and check its reported status. A schedule does not replace a required human approval.
02

GPU pool

Provide models and computing capacity to suit each task: on your own computers, on your own network or through external providers you choose to connect.

Pools and utilization
Group available model servers and review their reported capacity and utilization.
Own GPU/model servers
Connect your own computers and servers for AI processing in your environment.
External GPU/model providers
Connect additional computing capacity or models when needed. Required content may then be processed outside your installation.
Queue
See which model calls are waiting for available capacity. This technical queue complements the process job view.
Model and capacity limits
Limit permitted models and available capacity, such as the number of model calls processed at the same time.
Recent calls
Review recently reported model calls and their available run data.
03

Settings → Connections

Configure external services deliberately. Which connection a process uses and which data it needs are part of setting up that process.

Jira · optional
A Jira connection is planned under Settings → Connections to link c9n jobs with Jira issues when needed. Jira is not required for local processes. This does not promise automatic synchronization; connecting Jira exchanges the required data with the configured Jira instance.

How they fit together: A process defines the workflow. A job is the actual work within it. The GPU pool provides model capacity. People are responsible for the required approvals.

Development status: The new menu structure, process execution and integrations are being built. See the release notes for the published scope of each version. The Architecture & Privacy page explains how local processing and external connections work together.

PROCESS RUNS · APPROVALS · HISTORY

What is running? What needs your decision?

Process runs, agent steps and approvals form a shared view of the work. Each open item has a next step and an assigned person.

Illustrative process views with fictional content. These cards are not an interactive product screen.

01

Process run

Keep the context visible.

Customer requestAnalysis completeDraft ready
  • Agent steps: Fink · analysis and draft
  • Current step: waiting for Lea

Next step: obtain human approval.

02

My approvals

The process needs your judgment.

Review the replyAssigned to LeaApproval pending
  • Review the assignment and agent result
  • Approve, return for rework or reject

Next step: review the content and make an explicit decision.

03

History

Results and feedback in the workflow.

One revisionFeedback recordedReady for review
  • Lea: add the missing detail
  • Fink: revised the draft reply

Next step: review the new version again.

ARCHITECTURE & PRIVACY

Your processes contain
confidential information.

Documents, internal rules and human decisions belong to the same workflow. The architecture page explains why that information needs protection and how the planned local c9n installation works with supporting cloud services.

Fully local processing requires your own models and a workflow without external processing services. The shared package with Hermes and GPU management is being built.

Understand architecture and privacy →

ROLES IN THE SHARED PROCESS

Who builds? Who works? Who approves?

Roles follow responsibility in the process. People preparing agent steps need different tools from those reviewing and approving assigned results.

Illustrative role view · proposed navigation

Manage the team and setup.

Manage access and roles, and configure agent and model connections for the team.

Assign responsibilities

Who can build processes and who handles approvals?

Illustrative role view · proposed navigation

Build processes from steps.

Choose agents, define their assignments and add human approval steps with an assigned person to the workflow.

Prepare a request process

Analysis → draft reply → approval by Lea

Illustrative role view · proposed navigation

Review. Decide. Hand over.

Open an assigned approval, review the task and agent result, and explicitly approve, return for rework or reject it.

A draft reply is waiting for you

Only your approval allows the next defined step.

These views explain the tasks for each role. They do not change real permissions and are not a login.

The process sets the framework

Define the goal, sequence and permitted next steps during setup.

The agent handles its step

The assignment, inputs and expected result define the task. Passing a test does not replace a required human approval.

The person grants approval

Assign the step to a responsible person. Without their approval, the illustrated process must not continue past this point.

Scope: These role views show the intended initial scope. The new process engine and its access rules are being checked during development. A navigation illustration does not establish executable processes or configured permissions.

Discuss process setup and approvals →