v1.51.2Running only the tests that cover a change works in every test tool

A coding agent for the whole team, checked at every step.

A guard stops risky tool calls before they run. The team's Harness runs checks and has a subagent review the change. One dashboard for Terminal and the browser, with your model or the company's.

npm i -g @piagent/platform

Get started GitHub

MITNode.js ≥ 22.19macOS · Linux · Windows (WSL2)

Conversations

Fix money rounding

Paginate /orders

Read the project

Harness

check · review

fix/invoice-rounding

›

·Read src/billing/invoice.ts, tests/invoice.test.ts

✕Guard blocked: read .env · protected file

✎Edit src/billing/invoice.ts +12 −3

▶npm test — sandbox, no network 48 passed

◆Review · subagent read the diff: no blocking issue

✓ Check passed on the final code · Review found no blocking issue

The agent wants to run git push origin fix/invoice-rounding

Allow onceDeny

54/54tasks resolved in the production benchmarkcodex-cli, same model: 48/54

−61%fresh tokens on the same workload95% confidence: at least −51.6%

52/52secret patterns redacted from output0/30 ordinary texts redacted by mistake

142releases since 2026-07-214,400+ automated tests on macOS and Linux

Why it is different

The agent works fast. The platform keeps it right.

Every tool call goes through the guard, every risky action waits for you, and in company mode a turn is only reported done once checks ran on the final code.

A guard on every tool call

Fails closed: protected files, permission profiles, oversized files, edits on a stale read. When in doubt it blocks, it does not guess.

✕ read .env — protected path
✕ edit src/app.ts — read it again first
! git push — waits for one approval
✓ npm test — allowed

The team's Harness checks and reviews

Studio sets the plan, the checks and the review rounds for the whole team. A subagent reads the diff on its own; the outcome sits right under the answer.

plancheckreviewmain · research · review

One approval at a time

Push, publish and network commands run only after you press Allow once. There is no "always allow".

A sandbox in company mode

Seatbelt on macOS, bubblewrap on Linux and WSL2: no network, no Keychain, no .env or SSH keys.

The dashboard is an editor

Folder tree, editor, terminal, hunk-by-hunk review, a Timeline to restore turns and /undo. 13 colour themes.

Long tasks finish

A passing model error is retried; a Studio update is waited out, then work resumes. One turn can run for 24 hours.

Team memory

The agent remembers the team's decisions and rules per repository and branch; memory_ask answers with sources.

Front-end change reports

Before and after screenshots, each change in a numbered red frame, with checks and E2E. Shown right in Documents.

One turn

You say what you need. The rest is verified.

Ask

Plain language, no task contract or workflow command.

Read and edit

The agent reads the files it needs and edits with a clear diff.

Guard checks

Every tool call meets the policy; risky ones wait for you.

Check and review

Tests run on the final code; a subagent reviews on its own.

Report

One closing line: done, stopped part-way, or failed.

Real screens

Captured from a running dashboard

A company-mode turn: read, edit, run tests; the closing line says the check passed and review found no blocking issue — A company-mode turn: read, edit, run tests; the closing line records check and review.

Two modes

One dashboard, two ways to use models

Switching modes does not change how you work: the same guard, the same history across Terminal and the browser.

Personal

pi

Pick any model and thinking level
The same guard and protected paths as company mode
Push and publish still wait for approval
piagent dashboard opens the web interface

Company

piagent studio --web

Agent Watch keeps the company key in Keychain; Agent Studio picks the models and the Harness for the team.

The model never sees the key
Sandbox without network, opened one approval at a time
Checks and review run by the Harness
Usage, budgets and wrong-project warnings in Studio

Install

Running in two minutes

piagent-setup installs the exact Pi release this version pins, initializes the project and runs the doctor.

Install

Run it in the project folder.

bash

npm install -g @piagent/platform
piagent-setup

Open

Terminal, or the dashboard in your browser.

bash

piagent dashboard
pi                    # Terminal

Update

The dashboard tells you about new releases every hour.

bash

piagent-update

One PowerShell command

Installs WSL2, Ubuntu, Piagent and Agent Watch, and adds Piagent to the Start menu. Run it again to update.

PowerShell

irm https://raw.githubusercontent.com/Vt-mmm/agentwatch/main/windows/setup.ps1 | iex

Details

The first run restarts once to enable WSL2. The company connect code is optional.

Windows guide

Changelog

Steady releases, every change written down

Each release says what changed and which Studio or Agent Watch it needs. See the full changelog.

v1.51.2Nov 3, 2026

Running only the tests that cover a change works in every test tool

When you chose to run only the related tests, the harness took many single-test commands for the whole suite and refused them. It now reads one file, filter or module the same way in each tool.

A task left running overnight no longer waits for an answer until morning

In company mode the harness asked how much to test, or whether to push, and waited for you with no limit. A question nobody answers now goes away after ten minutes and the cheap answer goes on.

Follow every session from outside it, and a new Always bypass access

With many sessions open, you see which one is processing, paused or failed without opening it, from any terminal and on the Herdr sidebar. Always bypass, a new access you choose yourself, removes the sandbox and asks only about commits, pushes and removing a critical path.

Hand work to the agent. Keep the decisions.

Two commands to install. Docs in the order you need them.

Install Ecosystem Benchmark