1. Home
  2. Blog
  3. Workflow guide
Workflow guide

Build an AI agent from a plain-language description

Most agent builders hand you a form with twenty fields. In Shiva Agents you can skip all of them: type a name, describe the job the way you would brief a new colleague, and press Build and start. Shiva designs the agent, hires it into your company and starts its first task while you watch.

What you type

Open Agent Studio in the app. There are two fields:

  • Name — optional. Leave it blank and Shiva picks one.
  • Description — as long as you like. The more specific it is, the better the agent.

Under the description, small buttons add a starter line for the things people most often forget: Duties, How often, Who it serves, Tone and language, Good work looks like and It must never.

Agent Studio with the name Kavya and a description of a daily market briefing job
The whole form: a name and a description. This is the description used for the example below (a company name is pixelated).

What Shiva builds from it

A model reads the description and designs the full agent. Every field is then checked against what the platform actually allows, so the agent can only use real tools and sensible schedules:

  • Role, goal, background and working style, written from your words.
  • An operating manual in the agent's own memory: duties, process, quality bar and the things it must never do.
  • Facts worth remembering from your description (names, numbers, links, preferences), saved to its long-term memory.
  • Tools chosen from the platform's tools. Agents built this way never get tools that send data outside the platform or run code; you can grant those later on the agent's page.
  • Standing orders: one per recurring duty (up to five), each with its own schedule in IST, never more often than every 15 minutes.
  • Success measures, and a first task that produces something useful today.

Your full description is also kept word for word in the agent's core memory, so nothing you wrote is lost to summarising.

Progress card: reading the description, hiring Kavya, giving a first task, starts working
Building takes a few seconds to a minute, depending on the model.

What it does on day one

The agent is hired at autonomy level 3: it works end to end without waiting for you, within the tools it was given. It wakes every 15 minutes to pick up new tasks and any standing order that is due, and it starts its first task straight away. The Studio page streams the work live, with links to the full run, the agent's work list and a chat with it.

In our example, the description asked for a weekday gold-market briefing. Shiva hired Kavya, Market Intelligence Analyst, with two standing orders (a daily briefing on weekdays at 09:00 and a weekly summary on Fridays at 09:00), four success measures, ten tools and six remembered facts. Designing and hiring took about half a minute in our test run.

Kavya, Market Intelligence Analyst, autonomous level 3, standing orders and success measures, already working
The hired agent: autonomy, schedule, standing orders and success measures, with its first task already running.

Its first briefing arrived within a minute. It was also honest about a limit: the live price pages it tried were not accessible to automated reading, and it said so instead of inventing numbers.

The first task finished: a market briefing with sections What Changed and Why it Matters
The first task, done without anyone in the loop.

How to write a description that works

  1. Say what “done” looks like. “A briefing under 250 words with sources” beats “keep me updated”.
  2. Give the rhythm. “Every weekday at 9:00 IST” and “every Friday” become standing orders with real schedules.
  3. Name the audience and tone. “For me, plain English, numbers first” or “customers, polite Hindi and English”.
  4. Write the never-list. “Never give investment advice”, “never promise discounts”. These go into the operating manual.
  5. Leave out personal data. Describe the job, not your customers' phone numbers.

Limits worth knowing

  • An agent can only use the tools it has. Reading public web pages works; pages behind logins, paywalls or bot protection usually don't.
  • Agents built in one step cannot send data out or run code until you grant those tools on their page, where the approvals and autonomy settings also live.
  • Each agent has a daily token budget, and the kill switch stops every agent at once.
  • Prefer full control? Set every field yourself opens the step-by-step builder for persona, tools, autonomy and schedule.

Next: read how autonomous agents stay safe, or see a daily briefing agent in practice.