News · 2026-11-08

A developer’s switch to Pi puts the coding-agent harness in focus

Developer Caleb says he moved his primary coding workflow from Claude Code to the Pi coding-agent harness because he preferred its customization and smaller core. The decision highlights a choice about the software surrounding an AI model: convenience and built-in controls on one side, and more user-managed workflow on the other.

Key facts

• Caleb published his account on September 21, 2026, with an update on November 7.
• Pi's project site documents more than 15 model providers and mid-session model switching.
• Pi's maintainers describe a customizable core and omit several built-in features, including permission pop-ups.
• Pi's release history lists version 1.5.0 on November 2.

The model is only one part of a coding assistant. A harness is the surrounding software that gathers instructions, connects tools, manages context, asks for permissions and presents results. Changing harnesses can alter a developer's day without changing the underlying model at all. Caleb's post describes a harness switch; it does not say he stopped using Anthropic models.

His complaints were grounded in a particular workflow. He wanted project instructions shared through AGENTS.md, smoother use of selected editor text across sessions, and fewer changes to default instructions and tools. He also found some features and output organization poorly suited to how he reviewed work in Visual Studio Code. Those are first-person observations, not an independently controlled comparison of the products.

The timing of one grievance matters. Caleb notes that Anthropic announced support for AGENTS.md after he had decided to leave, and Anthropic's current memory documentation describes that support. Treating the earlier frustration as a permanent absence would mislead readers about the present product.

Pi's maintainers describe a system shaped through extensions, skills, prompts and themes. Imagine buying a workbench: a fully equipped bench comes with drawers and fixtures, while a modular one gives the owner more choice about the layout. The modular bench may fit a particular job better, but the owner has to install and maintain more of its equipment.

Pi's repository makes that trade-off explicit. Built-in plan mode, sub-agents, permission pop-ups and several other features are omitted from the core; users can add workflows with extensions or other tools. Pi runs with the launching user's permissions and does not provide its own permission system for filesystem, process, network or credential access. Its documentation says users who need stronger boundaries should use a sandbox. A smaller core is therefore a preference with operational consequences, not a guarantee of safer behavior.

Caleb says he built an extension for a multi-tab workflow in about 20 minutes and found the resulting setup smooth. That is a useful concrete example of the flexibility he values, not a promise that every user can reproduce it in the same time. Similarly, his counts of releases and prompt changes are his own tally; the dossier did not independently audit them. Pi itself continues to change: its release history documents a version after his switch. Customization does not mean software stands still.

In the Hacker News discussion, some users describe sticking with Pi for weeks or months. Others say they returned to Claude Code because subscription economics were better for their use, missed bundled features, or kept both tools. A defender of Claude Code argues its harness is tuned for Anthropic models, including context management and prompt caching. These are informative experiences and arguments, but self-selected comments do not measure the frequency of switching or establish which workflow performs better.

The strongest counterargument to the lean-harness pitch is that defaults save work and can provide useful guardrails. An agent with unrestricted access under the user's account needs boundaries outside the agent if a task demands them. Conversely, a bundled product's defaults may be inconvenient for a developer with an established review process. The right comparison includes setup, subscriptions or API charges, integrations and authority granted to tools—not just the model name or size of a system prompt.

The discussion is evidence of demand for control over the layer around a model. It is not evidence of a broad exodus. A representative comparison would need common tasks, complete costs, security settings and outcome measurements, none of which this small set of accounts supplies. Ground Truth's tool-use lesson explains why the harness matters when a model can act as well as answer.

Primary source, verified: read the paper →

Ground Truth. (2026, November 8). A developer’s switch to Pi puts the coding-agent harness in focus. Ground Truth. https://groundtruth.day/news/developer-chooses-pi-over-claude-code.html

@misc{groundtruth:developer-chooses-pi-over-claude-code,
  title  = {A developer’s switch to Pi puts the coding-agent harness in focus},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/developer-chooses-pi-over-claude-code.html}
}

Topics: coding-agents · developer-tools · pi · industry