# AI engineering, October 3, 2026

The Forward Pass daily issue, October 3, 2026. Source: https://theforwardpass.net/archive/daily/2026-10-03

> This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.

_A new GPT-6 model guide, Pi ships Pi Durable agent framework, Apple says it will tighten macOS Full Disk Access controls, GitHub opens Copilot code review APIs to Pro, Pro+, Max, Business and Enterprise_

One macOS permission can let an app read your files, mail, messages and browsing history. Apple says it will add controls to it, because desktop AI agents may reach personal data without users fully understanding it.

Meanwhile, a new guide for the GPT-6 family walks startups through model choice, reasoning effort, prompts, skills, tool coordination and production prep. Pi shipped Pi Durable, an experimental framework for long-running agent apps that you can install from npm today.

If you read one thing today: Pi Durable. It gives you restart-resumable agent runs, and its source is about 15,000 lines excluding tests, so you can study it or drop it into your own project this afternoon.

## GPT-6 family gets a startup guide covering model choice and reasoning effort

**Top News**

A new model guide for the GPT-6 family is written for startups. It covers the setup decisions a team faces when building on these models and preparing them for production.

Here's what it covers:
- Model choice, so you can match a GPT-6 family member to your workload.
- Reasoning effort, and how to set it for your tasks.
- Prompts and skills, for shaping how the model behaves.
- Tool coordination, for apps where the model works with several tools.
- Production workflow preparation, for moving a working build toward real use.

Why care? It puts the key GPT-6 setup choices for a startup team in one place.

Sources: [openai.com](https://openai.com/index/practical-guide-building-gpt-6)

## Pi ships Pi Durable, a roughly 15,000-line framework for agent apps

**Top News** · 492 HN points

Pi Durable is an experimental framework that makes agent apps resume where they left off. It shipped alongside Pi 1.0 and sits next to the Pi coding agent rather than replacing it.

Here's what changed:
- Work runs as checkpointed tasks, so after a restart the harness resumes anything unfinished.
- Interrupted model requests are retried, and interrupted tools rerun only when declared safe. Otherwise the model is told they were cut off.
- It handles concurrent conversations, forks, multiple clients, steerable runs, hooks and background compaction.
- Storage ships in memory, SQLite and JSONL flavors, and SQLite and JSONL can run on Bun or a Cloudflare Durable Object with a small adapter.
- Conversations store tool names, not code, so you can update implementations while the harness runs.

One catch: it is experimental, the API may change, and it only runs where there is a JavaScript runtime.

Try it: `npm install @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chord`

Sources: [earendil.com](https://earendil.com/posts/pi-durable/)

## Apple says it will add macOS Full Disk Access controls over AI agent risks

**Top News** · 165 HN points

Full Disk Access was built so backup apps could work. It can also let an app reach your files, mail, messages and browsing history. Amid concerns about desktop agents' access to personal data, Apple says it will add controls to this permission.

Here's what changed:
- Apple's stated safeguard is more explicit user action before an app can receive this level of access.
- The motivation is that desktop AI agents may access personal data without users fully understanding it.
- A journalist said Meta's Muse app on Mac knew the content of his private messages, a claim Meta disputed.
- A Wired report cited a flaw in ChatGPT's Mac app that could have exposed sensitive data.

One catch: Apple has given no release date or macOS version for the change.

Why care? If your agent depends on Full Disk Access, expect users to need to take more explicit action.

Sources: [techcrunch.com](https://techcrunch.com/2026/10/02/apple-says-its-tightening-macos-full-disk-access-controls-due-to-new-risks-from-ai-agents/)

## Signals

1. [Agent Error Dataset corrections lift verifier pass rates from 18.4% to 51.1% on 3,062 replay pairs](https://arxiv.org/abs/2609.40111)
2. [VoiceStudio runs voice cloning, dubbing and audiobook creation locally in 646 languages](https://github.com/debpalash/VoiceStudio) · 40 Stars today
3. [GitHub opens Copilot code review APIs to Pro, Pro+, Max, Business and Enterprise plans, with per-request effort levels](https://github.blog/changelog/2026-10-02-copilot-code-review-api-support-and-new-default-effort-level)
4. [Cloudflare AI Gateway adds web search API with Exa, Linkup and Ceramic.ai for live model context](https://blog.cloudflare.com/introducing-web-search-api/)
5. [PageIndex replaces vector databases with LLM reasoning over a document tree index for RAG](https://github.com/VectifyAI/PageIndex) · 13 Stars today
6. [llama.cpp adds decision models that score options in one forward pass, as fast as 3 ms](https://huggingface.co/blog/ggml-org/decision-models-in-llamacpp)