Reading the agent blueprint (watch, think, run, deliver) Guide

As you describe an agent to the AI builder, Praxivara draws a live blueprint of the agent it is assembling. The blueprint is your at-a-glance summary of how the agent behaves, organized around four moments: what it watches, how it thinks, what it runs, and what it delivers.

Reading it well helps you confirm the agent will do what you intended before you ever run it — and makes it obvious what to ask the builder to change.

Why the blueprint matters

The builder sets your agent up in plain English. The blueprint is how that conversation becomes something you can inspect. Each time you add a capability or refine the instructions, the blueprint updates so you can see the effect of your request immediately.

Tip: Treat the blueprint as a checklist. If any of the four moments looks incomplete or wrong, describe the fix to the builder rather than editing settings by hand — it will wire up the pieces and update the blueprint for you.

Watch — what starts the agent

The watch section shows what causes the agent to act. That might be a schedule (for example, every weekday morning), an incoming event from a connected service, or a request you send it directly.

When you read this section, confirm two things:

• The trigger matches how you actually want the agent to start.
• Anything it needs to watch a connected service is in place — an integration that appears here but is not yet connected cannot fire.

Think — how the agent reasons

The think section reflects the agent's purpose and the method it follows. This is where its instructions and any installed skills show up. A skill is a proven, step-by-step playbook for a specific job, so if your agent needs to follow a particular approach — writing an on-brand blog post, monitoring a brand — you should see that reflected here.

This is also where the agent's grounding lives. Knowledge you have given it (product docs, brand guidelines, price lists, FAQs) shapes how it reasons on every run. If the blueprint suggests the agent will answer too generically, it may be missing key knowledge.

Skills bring their own tools

Installing a skill also grants the extra tools that skill needs. So if you notice the agent has capabilities you did not add by name, a skill likely brought them along — that is expected.

Run — what the agent can actually do

The run section lists the agent's tools and integrations: the concrete actions it can take. Built-in tools are always available — things like emailing you, saving results, searching the web, working with files, and generating images. Integrations are the connected services it uses, such as Gmail, Slack, or Shopify, shown with the account each one uses.

Use this section to answer a simple question: does the agent have the means to finish the job? If it lacks a capability, ask the builder to add it (for example, "let it send WhatsApp messages"). If an integration appears but is not connected, connect the account so the agent can use it.

Note: An agent with no explicit tools set can still use all of the built-in tools. Integrations are the pieces that require connecting an account first.

Deliver — what you get back

The deliver section shows the agent's output and how it keeps you posted. Outputs might be reports, PDFs, or images, which you can preview and download later. Notifications control how the agent reaches you — an emailed summary, a message on a channel you use, or staying quiet until you check in.

Confirm the delivery matches where you actually work. If you want a summary in your inbox or a ping on a channel, describe that to the builder and watch the deliver section update.

Turning the blueprint into a working agent

1. Read the four sections in order — watch, think, run, deliver — and check each against your intent.
2. Where something is missing or wrong, describe the change to the builder in plain English.
3. Run a test to confirm the agent behaves as the blueprint suggests.
4. If a change did not land the way you wanted, use version history to roll back and try again.

Once all four moments read correctly and a test run looks right, your agent is ready to work on its own.