Agents

#Answering versus doing

A question gets an answer. A task needs actions: reading a file, calculating, writing the result and checking it. A reply that says “I created the file” isn’t evidence that anything happened. Praus agents act through tools that run in your application, so every step produces a real result you can inspect.

#How it works

Praus runs a task as a controlled loop. Praus decides the next step; your application or terminal performs it and reports back what actually happened.

1. 1Receive the task

   A task, and the tools explicitly made available for it.

2. 2Choose a tool

   Praus picks a tool and supplies structured arguments.

3. 3Validate

   The tool name, arguments and permissions are checked before anything runs.

4. 4Execute

   The tool runs in your application or terminal.

5. 5Return the result

   The real output goes back to Praus, not a description of it.

6. 6Repeat or stop

   Until the task completes, fails, is cancelled or reaches a configured limit.

7. 7Report

   Verified outputs and a record of every step taken.

#Availability

The standard text requests in the API reference do not execute tools. Selected API preview accounts receive the native tool-call contract and execution setup guide when access is enabled.

#Agent in Praus chat

You can also select Agent in Praus chat for research, comparisons, planning, reports and data tasks. Praus can search public sources, read discovered pages, calculate decimal formulas and create downloadable text, JSON, CSV, Word or Excel results in a private cloud workspace. It reads saved files back before returning them.

Attach up to three Word (.docx), Excel (.xlsx), text-based PDF or text files, up to 5 MB combined. Agent can read their contents and return completed copies while preserving the originals. Word completion supports paragraph edits, Excel completion supports cell values, and PDF completion supports existing fillable fields. Scanned PDFs, signatures, macros and arbitrary terminal commands are not supported. New Excel workbooks can contain checked values or editable formulas generated from validated calculations. Agent checks the saved formulas and initial values; it does not claim native Excel recalculation or recalculation of existing uploaded workbook formulas.

Agent task allowances depend on your account’s policy version. Pro is an individual plan; allowances apply to your account. Free and trial plans do not include Agent. Check Settings → Usage for your current allowance; the table below does not mean every policy is active or available to purchase.

One accepted task uses one Agent task allowance, including tasks that fail or are stopped. Its internal model and web steps do not use monthly or purchased messages. Processing, token, web and safety limits still apply to every step. Each account can run one task at a time, with up to 10 steps, 20 tool calls and four minutes per task. Purchased messages and Deep Research credits do not increase the Agent task allowance. Direct API calls retain their API accounting and limits.

Each task can make up to three public web lookups, reading up to two discovered pages per lookup. Results retain source links and distinguish page excerpts from search snippets. Agent cannot sign in to websites, submit forms or perform other website actions. Retrieval availability and processing allowances still apply.

Formula calculations support decimal arithmetic, division, powers and references to earlier results in a batch, including forecasts and sensitivity analysis. Agent separates sourced facts from modeling assumptions and may ask for essential missing inputs. Recent conversation excerpts help it interpret follow-ups; every task starts with a fresh workspace, so files from earlier tasks are not automatically available. A task can stop at its processing or step limit.

#Request access

Tell us about the tasks you want Praus to carry out and the tools involved. Contact us and choose API access.