AgentsEnd-to-End Workflows
Agent Types and Path Selection
Understand the runtime modes, then lock the path before configuration work.
Core runtime modes
| Mode | Where it runs | Typical fit |
|---|---|---|
| External HTTP | Your infrastructure | Existing hosted agents and custom runtime control |
| Code (sandbox) | Pipelines sandbox | Agents you want Pipelines to execute directly |
Sandbox execution paths
| Path | What runs | Typical fit |
|---|---|---|
| Python function | Python entrypoint called by the platform | Framework agents and proxy-first tools |
| Shell command | Command executed in a seeded workspace | Coding CLIs (Claude Code, Codex, Cursor, Aider) |
Patterns layered on top
- Multi-turn: session-based evaluation across turns, not a separate mode.
- Multi-agent: topology + attribution, not a separate mode.
- Browser automation: a capability pattern layered onto either runtime mode.
Choosing your path
Lock the path below before opening any setup form.
Step 1: Choose runtime ownership
- External HTTP: keep execution on your infrastructure.
- Sandbox code: run execution in Pipelines-managed sandboxes.
Step-by-step guides:
- External HTTP, end to end: Runbooks
- External HTTP registration flow: Register an external HTTP agent
- Sandbox registration flow: Runtime setup
Step 2: Choose execution style
If you picked sandbox code:
- Python entrypoint: framework-based agent code + proxy tool calls.
- Shell command: coding CLI workflows in seeded workspaces.
Step-by-step guides:
- Python wrapper + proxy wiring: Agent SDK (Python)
- Non-SDK runtime ports: Porting without the SDK
- Coding CLI setup and execution: Run a coding CLI agent
- Local bridge and smoke tests: Local development
Step 3: Layer testing patterns
- Add multi-turn when conversational memory and behavior matter.
- Add multi-agent when delegation and handoff quality matter.
Step-by-step guides:
- Multi-turn setup: Multi-turn testing
- Multi-agent topology and attribution: Multi-agent systems
- Seed structure used by both: Task seeding
Step 4: Plan observability
Pick where your team will debug failures and validate quality:
- Run inspection surfaces: Inspecting runs
- Live trace events: Trace events
- Failure class troubleshooting: Troubleshooting