Pipelines

A pipeline is an ordered set of stages. Each stage has an agent and requirements for advancement. Configure under Pipeline.

Stages

Per stage: name, agent, advance mode, requirements, pull request flag (on unless turned off).

New projects start with three stages (engineering requirements, implementation, code review), each with manual approval and each opening a pull request after a successful run. The Product Manager, Product Designer, and QA Engineer agents are still created; add a stage and assign one to bring it back.

A stage with an assigned agent starts when a card arrives. Sending a card back stops the agent and waits for user input.

Reordering: drag by the grip handle.

Hold conditions: agent failure or stop, failed requirement, or no assigned agent. After the last stage, the card is done (no further stage actions or agent runs). Reopen returns the card to the stage it finished in.

Early completion: drag to Completed or use Mark completed in the drawer. Intermediate stages are skipped. The card retains its current stage for Reopen (backlog if never started).

Gates

How do human gates work in a multi-agent pipeline?

In Bento, the agent pipeline at usebento.ai, every stage begins with a gate, and new projects start with three stages (engineering requirements, implementation, code review), each with manual approval. A person reviews the stage's output, then approves the card, sends it back, or steers the agent. Switch a stage to automatic once its requirements can decide: the card advances when every listed criterion passes, and holds when one fails.

All criteria on a stage must pass:

Re-evaluation triggers: run completion, GitHub webhook, manual re-check, or every five minutes.

Judge agent: second agent on the same card. Configure via skill and use a different model from the working agent. Incomplete verdict holds the card and displays the reason. New work triggers a new judgment.

Gates are part of every plan, including Free, which comes with 3 members and 5 agent hours a month. Create an account to set up a pipeline, or compare plans on the pricing page.

Agents

An agent is a harness and model pair. Create under Agents, assign to stages. Tool and credential reference: Coding agents.

Edit in place. Stages reference agents by id; delete-and-recreate leaves stages unassigned. Unsupported tool/model pairs are rejected at save time.

Deleting an agent removes its runs and transcripts. Card history is retained.

Skills: standing instructions included in every prompt. Define expected outputs, artifacts, and code changes per stage.

Export/import YAML from Agents or Settings, Config. See The agents file.

The pipeline file

Export/import under Pipeline.

Contents: stages, requirements, PR flags, named agents (model, skill), repository setup and test commands.

version: 1
pipeline:
  name: Default
  stages:
    - name: Code review
      slug: code-review
      description: Review the changes before they merge.
      gate: auto
      requirements:
        - type: checks_pass
      createPr: true
      agent: Code Reviewer
agents:
  - name: Code Reviewer
    tool: opencode
    model: openrouter/openai/gpt-5.6-sol
    skill: |
      Review the changes on this branch against what the earlier stages asked for.
repositories:
  - name: api
    setup: npm ci
    test: npm test

Import matches stages by slug and agents by name. Re-import updates in place and preserves card positions. A stage with no createPr in the file keeps its current setting, and a new one gets the default (on).

Stages omitted from the file are removed only when empty. Otherwise import fails and names the blocking stage. Repository commands apply to matching checkout names; unmatched names are reported.

CLI:

bento pipeline export team-pipeline.yaml
bento pipeline import team-pipeline.yaml --project "New service"

Settings, Config also exposes the agents file and pipeline file (select project for pipeline).

The agents file

Agent list without stages. Export/import from Agents or Settings, Config:

version: 1
agents:
  - name: Code Reviewer
    tool: opencode
    model: openrouter/openai/gpt-5.6-sol
    skill: |
      Review the changes on this branch against what the earlier stages asked for.

Import matches by name. Omitted agents are unchanged. File import does not confirm agent deletion side effects.

CLI:

bento agents export team-agents.yaml
bento agents import team-agents.yaml

Repository commands

Sandboxes include git and agent CLIs only. No language runtime. Configure under Repositories:

Setup runs once per card, not per run. Non-zero setup exit fails the run before the agent starts; output appears in the transcript.

The test command is passed to the agent for self-validation during the run.