Skip to main content

Project

Project bundles agents and steps to carry out work. In the Marketplace edition, you place and run agents registered by an administrator (registered agents) on the steps. In a workflow, each run is a case: it holds its settings and artifacts as versions, and a step where a person decides checks them before the work goes on.

The ideas are the same as in the SaaS edition. This section is written with the Marketplace edition's facts so that you can read it through in the order of how it works, quickstarts, guides, and references. Check the differences from the SaaS edition in the SaaS Edition Mapping Reference.

Main differences from the SaaS edition​

SaaS editionMarketplace edition
Agents placed on stepsUsers create them in the project (goal (instructions), guardrails, playbook, skills)Container agents an administrator registered in the admin console. They are not created in the project, and their contents are not changed there
What it takes to place one in a workflowIt has a goal (instructions)An administrator has registered a contract and it is certified
What you change per step(You fix the agent)The step settings (parameters, connectors, input bindings) / environment variables (cycles)
Hand-over between stepsThe output of the previous step passes to the next stepIt passes only when you bind the files
Running a workflow“Run now”“Start case”. Settings and artifacts are held as versions
Shape of a workflowIt can divide into lanes and run them at the same timeSerial. There are no parallel lanes and no “Run at the same time”. With a Branch step, the lane to go to can be chosen by the previous step's result (there is no “question” to check)
Review methodStop and wait for a decision / Notify and proceed after the deadlineStop and wait for a decision / External wait (a GitLab MR being merged). The deadline does not decide it
Ways to startManual, time, external API, emailManual, time, external API (no email)
Creating with AI, consulting on improvementsAvailableNot available
Project memoryAvailableNot available
Execution rulesThe basic rules and forbidden operations act on the runThe basic rules decide whether it can start (workflows). Outbound traffic is restricted with NetworkPolicy and “Banned URL” in the admin console

How a project is built​

A project has steps, and you place registered agents on those steps. What a registered agent does is settled by its container image (and, in a workflow, by the contract an administrator registered). On a step where a person decides, no agent runs: a reviewer looks at the case's artifacts and settings and decides.

CaseAgentStepReviewDone
TermMeaning
Registered agentAn agent an administrator registered under “Marketplace AI Agents” in the admin console. It runs as a container image built with the SDK
ContractThe agreement on the parameters, files, artifacts and connections a registered agent receives, its effect on external systems and its execution limit. Only agents used in workflows have one
CertificationThe contract has, all in place, the SDK version, adapter version, image digest, test results and supported execution rules confirmed by testing. Only certified agents can be placed in a workflow
Step settingsThe parameters, connectors and input bindings set for each placement in a workflow, within the range of the contract
CaseOne run of a workflow with registered agents placed. It holds its settings and artifacts as versions
External waitA review method that moves on to the next step according to the state of an external service (a GitLab MR being merged)

The other terms (step, placement, review step, run record, execution rules, ledger) are the same as in the SaaS edition (How a project is built).

The two types​

You choose the type when you create the project. The default in the Marketplace edition is workflow.

CycleWorkflow
How it startsYou set a period for each registered agent placed, and it runs on that period. It can also be run by webhookStarted manually (Start a case). It can also be started at a set time or from the external API
How it proceedsRegistered agents do not wait for one another; each runs on its own periodOne execution (a case) goes through the steps in order. The next step starts after the previous one finishes. A Branch step chooses the lane to pass through
How it endsIt does not end. It keeps running until you stop itEach item ends (completed, rejected, failed, or cancelled)
What can be placed on stepsRegistered agents used in chat (no contract or certification needed). A review step cannot be placedCertified registered agents. Review steps and Branch steps can be placed
Example of suitable workCheck a state at fixed intervals and notify if something is wrongReceive one document, process it, and finish it through a reviewer's check

How you use it​

1Administrator

Register the agent

Register it under Marketplace AI Agents, and add a contract and certification if it is used in workflows.

2Operations

Allow traffic

If the agent uses external services, allow the destinations.

3Project

Place it on steps

Place registered agents and enter the step settings.

4Project

Take cases forward

Start a case, and move it on at review steps after looking at the versions.

5Project

Look at the record

Choose what to fix from the run records and metrics.

Where to read at each stage.

Where things are in the UI​

The project screen shows a row of tabs. Because “Workflow history” exists only for workflows, a workflow shows five tabs and a cycle shows four (a person who opens it only with an organization operations role does not see the Settings tab). The editing screen opens with the “Edit” button at the top, not a tab (shown to owners only). There is no “Library” tab.

TabWhat you see
OverviewFrom the top: “Project overview” (a description written by AI from the purpose, steps and Agent settings, or the project description; not shown when there is neither), “Results of this work” for the period you choose (a sentence on the results and the indicators), a diagram of how the registered agents placed in the steps relate (its right-hand column switches between “Basic info”, “Metrics” and “Latest artifacts” for each step; “Latest artifacts” is not shown to people who open the project only through the organization operations role), for a cycle the band of the last 24 hours for each agent, and at the bottom “Project info” (a one-line summary; “Show All” opens all of it). Runs whose completion is not measured (“Run ended (completion not measured)”) appear separately in the legend
Execution historyThe list of executions. Open one to see its execution record. The record of operations inside the project (Records (audit trail)) is also here
Workflow historyThe most recent cases, including finished ones. “Case details” opens the case screen
Action neededA list of things that need a person. There are eight kinds: awaiting approval, overdue, waiting on external, failed, awaiting rework, sweep failed, past deadline (run), and stalled (Record Guide)
SettingsBasic information, Triggers and notifications, Connectors (MCP, A2A), Members, and Execution rules. A workflow also has “External access (A2A, MCP)”

In the Marketplace edition, the following differ.

WhereDifference
Top of the project listThere is no “Agents” tab (Home, Projects, Automations, Insights)
“Add Agent” in the editorThe list switches to the list of registered agents
Top of the project screenIn a workflow with registered agents placed, “Start case” instead of “Run now”
Workflow history, run record“Case details” leads to the case screen (progress, settings, files, records)
Settings > Triggers and notificationsThere is no starting by email
Settings > ConnectorsThere is an “External-wait webhook secret” (workflows)
Settings > External access (A2A, MCP)In a workflow, once the owner publishes it, external agents (A2A) and MCP clients can start a case with a request, in environments where an administrator has opened the entrance (MCP, A2A)

Projects can also be used through the external API. It can be called with a service account token (System Integration API Reference) and with the token of a user who signed in from an external tool (User API Reference (Projects)).

Use It from an AI Client​

Install the AGENTIC STAR plugin in Claude Code or Codex, and you can work with projects just by asking. For example:

  • "Create a workflow that uses the registered agent 'Contract Check' to check incoming contracts and ends after I review them"
  • "Find out why recent runs did not go well, and tell me who should fix them"
  • "Go through this morning's approval requests for me, checking what each contains"
  • "Show me the success rate of each project for the last 30 days"

For how to install the plugin and connect, see AI Client Connection Guide (MCP) — Install the Plugin. The AI can only do what the connected account or service account is allowed to do.

How It Works​

This gathers what the screens alone do not make clear. Read it before you start setting up.

  • Projects and Steps — A step is where registered agents are placed. The previous step's output is passed only when you connect files
  • Cycles and Workflows — A workflow ends as a case; a cycle keeps running placement by placement
  • Who Can Decide What — Even an owner cannot decide a review unless named. Only the owner can correct a case's settings
  • Human Review — A review step decides whether to move the case on. The deadline does not decide
  • How Much to Delegate — Run rules are a check before launch. Operations during a run are not stopped by the platform
  • Records and Metrics — An execution record is not a report card. “Completion not measured” is neither success nor failure

Quickstart​

Guides​

  • Design Guide — What to decide as a business matter, how to divide it, and what to hand to steps and how far
  • Agent Guide — the agents you can place, step settings and environment variables, and what the agent receives
  • Case Guide — starting a case, versions of settings and artifacts, and cancelling
  • Review Guide — the deadline does not decide, and external waits
  • Trigger Guide — how starting differs, cycle sweeps and webhooks, and when it does not go to plan
  • Execution Rules and Traffic Guide — where execution rules apply, and how to restrict outbound traffic
  • Connection Guide — what to connect to, and where to notify
  • Record Guide — action needed, the runtime image, unmeasured results, and why a run could not be launched

References​