TypeSafe: Jev is not a decision model
TypeSafe AI spent the week arguing its own model should never be called a decision model, because reliability comes from code composing judgments. Meanwhile Mike Taylor put a number on decision clones: 400 shots to reach 95%.
Built with it
ILIAS ISM puts a two-step spend gate on five write-enabled MCP servers
ILIAS ISM upgraded all five of his products to remote MCP servers (tool protocol for agents) that write as well as read, after noticing most only wrap GET requests and dump 20,000 tokens of JSON into the chat. The backlink tool previews the exact quote and only charges the saved card when the agent passes confirmPaymentSpend: true.
Milind S has a cloud agent opening bug fixes before the team reads the email
Milind S says users report bugs inside their chat bot, which asks a follow-up or two and sends a report with context, and a cloud agent usually opens the fix first. The agent runs 24/7, tests the build on a boat.dev machine and attaches a screenshot from the virtual machine to the pull request, which the team then reviews and merges.
Writing criteria
TypeSafe argues Jev makes judgments, and your code makes the decisions
The week's most-copied framing came from Allie the Icon, who argued that treating Jev's questions as decisions puts it at the wrong layer of abstraction and costs you reliability. Diogo Almeida said the biggest anti-pattern he sees is plugging Jev in to make big decisions instead of letting code do what code is best at.
reliability requires code, and so reliable decisions are made by composing reliable judgments in code
Moritz Kremb writes skill feedback into one doc the skill rereads each run
Moritz Kremb says he now builds most of his agent skills with a built-in feedback loop: a single document that is read every time the skill runs. Whenever he has feedback on an output, it gets added to that doc, so the skill improves without touching the skill itself. He calls it the single biggest thing you can do to improve agent skills.
The case against
Max Blade says every current agent harness is built wrong
Max Blade argues the "new thread" button is the problem, because it forces one human prompting one agent at a time. His alternative: the harness should spot recurring tasks, suggest automations, track results and modify workflows itself. He posted no implementation and no numbers.
the harness should actively try to REMOVE the human from the loop.
Also on the timeline
Mike Taylor: 400 shots to get a decision clone to 95%
In testing on Dan, Mike Taylor reports 60% accuracy one shot and 85% at ten shot. Reaching 95% took 400 shots.
Allie the Icon's shopping cart pattern handles multi-intent queries
Allie the Icon's shopping cart pattern recursively picks the next thing to do until Jev says it is done.
Frequently asked questions
What is an MCP server and why does it matter for spending?
An MCP server is a tool protocol for agents that can both read and write data. ILIAS ISM added a confirmation step called confirmPaymentSpend that requires a second call before charging a card, which is cheaper and safer than relying on the model to be careful with money.
How can agents help fix bugs faster?
Milind S describes a workflow where a cloud agent runs 24/7, tests bug fixes on a virtual machine, and attaches screenshots to pull requests for the team to review, often opening fixes before the team reads the initial bug report.
What's the difference between Jev making judgments and code making decisions?
TypeSafe argues that using Jev's questions as final decisions puts responsibility at the wrong layer and costs reliability, and that code should be what makes the actual decisions while Jev handles other tasks.
How can you improve agent skills without editing the skill itself?
Moritz Kremb builds agent skills with a feedback document that the skill reads every time it runs, so new feedback can be added to that document to improve performance without changing the skill code.
How many attempts does it take to clone a decision with high accuracy?
Mike Taylor reports reaching 60% accuracy with one attempt and 85% with ten attempts, but achieving 95% accuracy required 400 shots.

