fullauto.online

Harnesses

Pi 1.0 vs DeepSeek Harness v0.2: The Minimalist vs Plugin Harness Split — fullauto.online

In October 2026, two agent harnesses took opposite paths: Earendil’s Pi 1.0 doubled down on minimalism and transparency, while DeepSeek Harness v0.2 embraced an everything-is-a-plugin model. What the split means for builders.

Published
02 Oct 2026
Reading
11 min
Class
harnesses

half-life 120dfrom 02 Oct 2026

On 1 October 2026, Earendil shipped Pi 1.0 — the first stable release of its MIT-licensed agent harness — after months of community feedback. Four days later, on 2 October, DeepSeek announced Harness v0.2 preview with desktop installers for macOS and Windows. The releases highlight a growing divergence in agent harness philosophy: Pi leans into radical transparency and a fixed, auditable core, while DeepSeek embraces an everything-is-a-plugin model where even the agent loop itself can be swapped out.

Pi 1.0: Hardened minimalism

Pi 1.0 remains true to its original promise: four tools (read, write, edit, bash), a 150-word system prompt, and no hidden planner or subagents by default. What changed in the stabilisation round is the addition of Codemode — a JavaScript sandbox that enables native MCP support, tool composition, and access to non-LLM models like Jev (classifier) and image models — without adding opacity. The harness now also supports:

  • Extension support for virtual models
  • Deferred tool loading
  • Cache warming for Anthropic models
  • Mid-conversation system messages (transcript-aware)
  • A new full-screen TUI theme enabled by default

Alongside Pi 1.0, Earendil shipped the experimental Pi Durable package for long-running agent applications. Pi Durable adds checkpointing, forked conversations, and multiple steering clients while keeping the core loop inspectable. Both are MIT licensed; installation is a one-line curl or PowerShell script from pi.dev.

Earendil’s stance is clear: “We wait until something has proven itself, and only then do we consider adopting it; weighing its true functionality against its inherent added complexity.” The Pi team treats the harness as a transparent substrate — you can fork it, change at the source, and see exactly what runs.

DeepSeek Harness v0.2: Everything is a plugin

DeepSeek Harness (DSH) v0.2 preview, released 29 September 2026, builds on the Cordis meta-framework where models, tools, skills, sessions, sandboxes, storage, loops, and the UI all sit behind plugin boundaries. Any of them can be selected, swapped, or extended in configuration without touching Harness source code. The v0.2 release focuses on three areas:

  • Plugin management: A dedicated GUI page (macOS/Windows) lets users install plugins via npm name, disable, uninstall, and inspect each plugin’s description and source.
  • Automation tasks: A built-in plugin converts repetitive work into scheduled tasks with run history, adjustable frequency, and editable instructions — tasks continue running even after the application is closed.
  • Reworked file handling: Substantially changes how files, content previews, and code changes are displayed; users can control how much of the agent’s working process is visible.

DeepSeek reports roughly 60% of Harness users run third-party plugins, and the company plans an official plugin marketplace alongside a more stable plugin API. The harness can run as a desktop application or as a web interface launched from code (requiring only Node.js). Crucially, DeepSeek’s documentation warns of “compatibility-breaking changes” and notes the product page says core plugins and APIs will continue to evolve.

The split: What it means for builders

The Pi vs DeepSeek contrast reflects a broader tension in agent engineering:

  • Trust surface: Pi gives you a loop you can fully inspect, fork, and change at the source. DeepSeek trades some of that visibility for flexibility — you can swap the agent loop itself via plugin, but must trust the plugin boundaries.
  • Update cadence: Pi’s changes are additive and proven; DeepSeek’s preview status means APIs may shift, creating maintenance overhead for teams building on top.
  • Tool composition: Pi’s Codemode sandbox lets you combine tool calls in JavaScript without leaving the harness side. DeepShot’s plugin model achieves similar ends but requires crossing plugin boundaries to share state.
  • Long-running agents: Pi addresses durability via the separate Pi Durable package; DeepSeek bakes session persistence, autorestart, and multi-client steering into the core.

For teams that need to audit or modify the harness itself (e.g., regulated environments, security-conscious builds), Pi’s transparency offers a clear advantage. For teams that want to experiment with different agent loops, toolsets, or UIs without forking, DeepSeek’s plugin model lowers the barrier to swap components.

Both approaches acknowledge the same truth: the harness is becoming the strategic layer where agent safety, performance, and extensibility are decided. The model matters less than what you let it do, and how you can see and control what it does.

The harness split arrives amid broader agentic shifts:

  • Antigravity hard-shutdown: Google’s antigravity-preview-05-2026 harness stops accepting requests on 5 October 2026, pushing users to antigravity-preview-09-2026 which introduces PascalCase tool parameters and line-range file edits.
  • Devin Desktop: On 2 October 2026, Windsurf rebranded as Devin Desktop, expanding into an agent command centre for managing local and cloud agents.
  • Manus 2.0 and Cue: Launched 28 September 2026, Manus rebuilt its agent on a new harness (Cascade) and released Cue, a personal-agent app where each agent gets its own email, phone number, wallet, and computer.

These movements reinforce that the harness — not the model — is where differentiation, risk, and control now live in the agent stack.