---
url: https://pencel.ai/docs/guides/run-and-build.md
description: >-
  How to use Pencel's two composer verbs — Run to do work now, live and
  conversationally, and Build to set up a reusable, schedulable process — and
  when to reach for each.
---

# Run vs. Build

The composer on **Home** has two verbs: **Run** and **Build**. They share one input box, and choosing between them is the main decision you make every time you start something in Pencel.

* **Run** — do the work *now*, live and conversationally.
* **Build** — set up a *reusable, schedulable* process you'll run again.

This guide covers how to work in each and when to use which. For the concept behind it, see [Run mode](/concepts/chat).

## Run: do the work now

Run mode opens a live session with an agent that can use its full toolkit to get something done immediately — answer a question, produce a document, pull data, or kick off a workflow.

### 1. Pick who runs it

In the composer controls, choose an agent — or leave it on **Auto** and let Pencel route your request to a suitable one. Each agent brings its own role, guidelines, knowledge, and tools. You can also pick a **model**, attach a file, or dictate with voice.

### 2. Type what you want and press Run

Describe your task in plain language and press **Run**. The agent responds in real time — its reasoning and tool calls stream in as it works. Follow up, refine, or change direction without re-explaining; the session remembers the conversation.

### 3. Review what it did

Expand any assistant message to see its **execution trace** — which tools the agent called, the inputs and results, and its reasoning. This is how you build trust in what the agent produced.

::: warning Oversight
When a task would send or change something, the agent pauses and asks for **approval** first — based on your [guidelines](/concepts/guidelines) and the agent's plan-approval setting. See [Approval workflows](/privacy-and-security/approval-workflows).
:::

Each Run is a separate **session**, saved automatically. Start a new session for a new topic so unrelated earlier context doesn't bleed in, and revisit past sessions any time.

## Build: set up a reusable process

When you catch yourself doing the same thing again, switch to **Build**. Instead of doing the work once, Build helps you *construct* a repeatable process.

### 1. Describe the process

In Build mode, describe the outcome you want to repeat — for example, "every Monday at 8am, summarize last week's sales and post the highlights to Slack."

### 2. Answer a few setup questions

Build hands off to a guided setup where Pencel turns your description into concrete steps, asks any clarifying questions it needs, and lets you choose whether it runs **once** or on a **schedule**. Under the hood, Pencel compiles your intent into a plan with guardrails — see [Agentic compiler & runtime](/concepts/compiler-runtime).

### 3. Review and save

Review the proposed steps and approval points, then save. The result is a [Workflow](/concepts/playbooks) that shows up as a [Job](/concepts/jobs) you can run on demand or let its schedule fire.

## Which one should I use?

| | **Run** | **Build** |
|-|---------|-----------|
| **Goal** | Get this done now | Set it up to repeat |
| **Feels like** | A conversation | A short setup wizard |
| **Result** | An answer, an artifact, an action taken | A saved, schedulable [Workflow](/concepts/playbooks) |
| **Best for** | One-off tasks, exploring, ad-hoc work | Recurring reports, monitoring, briefings |

::: tip
Use **Run** to figure out the right approach, then **Build** to make it repeatable. Explore live, then capture what worked.
:::

To run a process you've **already built**, don't use Build again — start it from the [Jobs](/concepts/jobs) tab or let its [schedule](/concepts/triggers) fire.

## Quick forms with `/`

In either mode, type `/` in the composer to summon a pre-built form — no AI round-trip, the same panel every time. `/todo` adds a to-do; installed [Marketplace](/concepts/solutions) packs can add more. See [Tips & shortcuts](/reference/tips).

## Tips

* **Be specific.** "What were Q3 sales by region vs. Q2?" beats "What are our numbers?"
* **Set the format.** Ask for a table, bullet list, or a specific structure if you want one.
* **Explore in Run, commit in Build.** Prototype live, then save the approach as a Workflow.
* **Review traces.** Even good results are worth a glance — it's how you catch subtle issues and learn to trust automation.
* **New session per topic.** A fresh session gives the agent a clean slate.

## Next steps

* [Run mode](/concepts/chat) — the concept behind the Run verb
* [Building a Workflow](/guides/building-a-playbook) — go deeper on what Build produces
* [Creating Your First Agent](/guides/creating-your-first-agent) — set up a specialized agent to run with
* [Monitoring Jobs](/guides/monitoring-runs) — track what agents do during runs