AI Engineer building agent tools and workflows @WorkOS https://nitter.gq/t.co/FyPcujTvkz
Idaho
Joined December 2024
- Tweets765
- Following62
- Followers700
- Likes2.7K
Mitch Fultz retweeted
Rephrase it. "Do what a graphic designer would do when consulted to give our app a visual refresh. Greenfield approach."
Mitch Fultz retweeted
WorkOS is on the Fortune Cyber 60 list of most innovative cybersecurity startups.
The CISO survey published with it: 77% of enterprises run AI agents, only 34% give them their own identity.
Thank you @FortuneMagazine, @lightspeedvp, @AWSstartups. workos.com/blog/workos-cyber…
Mitch Fultz retweeted
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
Mitch Fultz retweeted
Based on my experience, yes, Opus taking shortcuts is highly dependent on the effort level. Low, medium, high are great as sub-agents, but for the main agent I cant trust less than xhigh.
Mitch Fultz retweeted
If you have a Droid Turbo you can fully unlock the bootloader btw github.com/fitchmultz/quark-…
Mitch Fultz retweeted
"Our initial evaluation suggested that GPT-6-Astra's score was in the late eighties followed closely by Opus 5.5. But, as we started to investigate the mistakes that the models might have made, the models, especially Astra, insisted that they were right and that it was the expert-supplied answers that were erroneous. When we went back to the experts it turned out, in multiple instances, that that the experts found some subtle errors buried deep inside their calculations. As they revised their answers, Astra's score started to inch up and it finally ended up with 98% (49/50)."
Mitch Fultz retweeted
Been to a bunch of @WorkOS events and honestly the merch and food were always what I remembered most 😂
But the Airlock live demo at init() today was really good. At first I thought it was just another auto mode for agents, then they showed how you can write policies in plain English to decide which tool calls can go through and which need a teammate to approve.
I was working on almost the same thing for Nuphos’ auto mode two months ago, before we pivoted.
They’ve taken it a lot further though, especially for enterprise teams. Really cool to see something I’d been thinking about work this well in a live demo.
Would love to integrate it into @NuphosAI. Agent governance is still a piece we’re missing.
Mitch Fultz retweeted
snyone claiming that GPT Sol 6.1 is not a great model has severe skill issues
stop running benchmarks and maybe, just maybe, use the model to get some fucking work done for once
then report back
Mitch Fultz retweeted
I built pi-posthorse, a Pi extension inspired by Codex's new approach to context management. Instead of relying on compaction, agents can inspect context usage, reset their context window, write durable notes, and search their full session history.
github.com/fitchmultz/pi-pos…
Mitch Fultz retweeted
This has been working amazingly well. Let the model decide what to keep in context and when to reset for a fresh window. No more lossy compaction summaries. github.com/fitchmultz/pi-hea…
Mitch Fultz retweeted
Introduction WorkOS Airlock: the permissions and control layer for AI agents
Mitch Fultz retweeted
Still keeping this lower context size indefinitely, even after the @thsottiaux tweet. Quality degrades, even for gpt-5.6, at larger context windows. Compaction in pi has been solid.
Morning. The last 48 hours of Codex and ChatGPT Work have been intense! Three important updates:
- Temporarily removing the 5 hour usage limit restriction for all Plus, Business and Pro plans
- Rolling out changes that will make GPT 5.6 Sol more efficient across the board and that will be reflected in less usage being used so that it can take you further. Exact impact to be quantified and shared
- We hit 6M active users, and are landing a usage reset in the next hour
Go do things
Mitch Fultz retweeted
Found in the app bundled agent.v1 protocol descriptors, not public docs.
Current Cursor Desktop/CLI builds include:
ToolCall fields 61–67: pi_read_tool_call … pi_ls_tool_call
ExecServerMessage fields 45–51: pi_read_args … pi_ls_args
ExecClientMessage fields 46–52: pi_read_result … pi_ls_result
And @cursor/[email protected] contains local LocalPi*Executor handlers for read/bash/edit/write/grep/find/ls.
Caveat: normal SDK/CLI runs still emit readToolCall / readArgs; I haven’t proven the selector that activates Pi mode yet.
Mitch Fultz retweeted
OpenAI-compatible endpoints get you access to Cursor models.
pi-cursor-sdk keeps Cursor’s agent loop as the runtime.
That’s the distinction: access vs integration.
Mitch Fultz retweeted
Your coding agent spawns a sub-agent.
Sub-agent runs 8 minutes. Main agent wakes up —
cache is cold. You just paid full input token price. 🤑
The issue: OpenAI's in-memory cache TTL is 5–10 min.
Sometimes 1 hour if GPU memory isn't busy.
Unreliable for any sub-agent workflow.
The fix: `prompt_cache_retention: "24h"`
Offloads KV cache.
24-hour deterministic TTL. No extra cost.
For pi, set env var to `PI_CACHE_RETENTION=long`
nitter.gq/i/status/2060843228618…
Mitch Fultz retweeted
Use the included /create-goal prompt template instead of writing a goal by hand. Agents write better goal completion contracts than humans do because they can expand a plain task into outcome, verification, constraints, iteration, audit, and blocked-stop requirements before calling the create_goal tool.
`/create-goal insert task and requirements here`
The template follows the Codex goal-writing practices from:
developers.openai.com/codex/…
developers.openai.com/cookbo…
Mitch Fultz retweeted
I used the Pi coding agent with the DeepSeek v4 Flask model to scrape 285K URLs from a website. It ran for 1.5 hours and costed me just $1 via the official DeepSeek API.
I used `pi-codex-goal` extension by Mitch for this. It's an amazing one.
If pi /goal ever retried itself into oblivion after a provider hiccup — v0.1.13 is for you.
pi-codex-goal v0.1.13 🚀
• provider errors back off + pause cleanly (/goal resume) — no more infinite hidden retries
• context overflow → pi auto-compact → active goal resumes instead of dying mid-run
• coalesced goal state — footer stays current, session log stops flooding
pi install npm:pi-codex-goal
What's your best pi /goal success story so far?
github.com/fitchmultz/pi-cod…
