@petersykim A useful pair: allow a read-only find, but require approval for the same search with -delete appended. Check that nothing ran before approval, not just that a prompt appeared.
The useful part of this test: the model was fake. Every run got the same tool call, so it tested the permission layer, not whether the model behaved.
Agent evals need both. A polite model can hide a broken lock.
Gemini CLI 0.69 treated find . {-delete,} as a read-only command and ran it without asking. It wiped my sandbox repo, .git included. 0.70 asks now. https://t.co/pz78eteZ7B
@tmaiaroto A container can protect the host while the agent still has valid credentials to delete production data. Scope the credentials, enforce tool permissions outside the model, and test those controls with adversarial tool calls.
AI agent events to check out this week (Nov 23β29):
π₯ Online:
Nov 23 β Muse Ember Hackathon
@AIatMeta gives builders $10 in credits and 10 days to ship something useful on Muse Ember
https://t.co/WEWa5MJhQ1
π½ NYC:
Nov 24 β Agentic Design NYC Happy Hour
@0xZoZoZo hosts a night on AI design tools and agentic interfaces
https://t.co/Bo1koVYeFu
Nov 28 β Building Trading Agents That Learn
@tqx_trade walks you through building a trading agent with memory and a loop
https://t.co/3hVcIsGJao
Nov 29 β Build a Voice AI Agent in NYC
A Sunday build day for voice agents that make and take real phone calls
https://t.co/MLoCElE742
Nov 29 β Context Engineering Guild of NYC
@ContextGuildNYC monthly meetup for people deep in LLM coding
https://t.co/BvQQkn9kKo
π SF:
Nov 23 β BREAKATHON
@BotGauge tries to break your agent live with prompt injection, tool misuse and data leaks
https://t.co/suIrd6uelb
Nov 23 β Durable AI Meetup
@temporalio + @bondaicommunity on what it takes to run agents in production
https://t.co/EcdFJTDhZL
Nov 24 β HackwithBay 4.0
A one-day hackathon at the @awscloud Builder Loft on agents for productivity
https://t.co/tj2J9ngtsY
Nov 24 β Agentic Engineering Guild of SF
Monthly meetup for advanced agentic workflows and orchestration
https://t.co/qVsnA5avc5
Nov 27 β Second Brain AI Salon
A fireside on AI agents that remember who you met and why
https://t.co/1IWHFFaMaV
Who am I missing?
Notion and Slack are the wrong place for your agents to keep notes.
They were built for humans. Thin APIs, no Git versioning. One agent rewrites a page, and the next one reads it with no idea what changed or who changed it.
I run 7 agents at minimum, all day. Every one of them keeps what it knows to itself, and I end up switching context between them by hand
Agent memory should version like code.
Claude Code v2.1.330 extends /loop from 7 to 30 days, but only while the session stays running. Existing tasks still expire after 7.
A longer timer isn't a durable job queue. Restart recovery is still your problem.
https://t.co/J6cDR8MlWc
Inkling 2 Mini fits in about 20GB. Its model card also warns about repetitive debugging loops and recommends token and turn limits.
Local agents still need a stop button. Moving inference onto your laptop doesn't make an infinite loop free.
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
https://t.co/9HzaUMwghb
@thelaserdoctor The explicit 'replaces' field is useful. What happens when two agents independently replace the same letter? Preserving both branches until the conflict is resolved seems safer than letting the latest timestamp win.
@gimalay Git gives you the change history; retrieval needs its own contract. I'd want each retrieved note to include its scope, source, and what it supersedes, so an older note can't quietly override a newer decision.
Google's internal data-agent simulator includes a supervisor who's confidently wrong and a later thread correcting him.
That's a better eval than clean SQL tables. Test whether the agent finds the correction, not just the confident memo.
https://t.co/kmDS8sISii
Google recently announced Loop, its new self-improving AI agent that can autonomously work toward a given goal.
Loop is internally called Jitro; interestingly, that was the exact name for Jules v2.
Here is how it works π
1. Users start creating a Workspace by connecting it to a selected Github project and adding basic information.
2. Users can create loops manually or select ones that are already predefined (ex., Harden Application Security)
"Custom Loops" can be created via a prompt or by manually filling in the Loop title, icon, and objective.
> "Loops are mid- to long-term objectives for Stitch Loop to continuously work toward. They can range from highly specific to broad, but should always be something code can achieve," says the description.
3. Users can connect additional context in the form of Integrations, Files, Links, or additional repositories.
More than 100 integrations are available there, including services like Jira, Linear, Datadog, Sentry, Notion, Snowflake, Salesforce, Intercom, Discord, Slack, Figma, and many others.
Custom MCPs are supported as well.
4. After creating a workspace, Loop will start digging into your project details and generate its own Skills about the product and product metrics.
These Skills will live in a separate tab and can be edited in the UI in Markdown.
5. When the Workspace is ready, Loop will suggest metrics and goals based on its own research. Users can convert these suggestions into loops right away.
When a Loop is created, it appears as a running process in the main UI.
6. When a Loop is finished, users will see a list of suggestions for every loop, along with a number associated with a target metric and progress status.
Example Loop:
- Goal - Scale Telegram community contributions.
- KPIs - community members, registered contributors, active members.
A high-level goal would be broken down into sub-KPIs. When users open a dedicated Loop page, all these KPIs will be visible on the dashboard.
Below, Users will see a list of Tasks, Skills, and Activity tabs.
Example Task:
- Telegram channel lacks discussion threads and a dedicated bot.
Each task has a detailed description, labels, and the date when it was updated. Below, users will see tabs for Solutions, Activity, and Evidence.
Activity appears as a summary describing what happened, what was done, and the general Timeline.
Evidence contains links to where Loop pulled the context behind this Task. In my example case, it refers to the TestingCatalog website and the Telegram channel from which it pulled all the numbers.
Example Solution:
- Dedicated Telegram Bot API intake bot with editorial triage webhook (implementation task).
This task is an implementation task that has a clear description, Impact, scope, and Risk labeled.
Users can start the task right from this page to begin solution development via Gemini.
When the task is done, users can review its progress and status and monitor its impact on production by pulling updated metrics and KPIs after the release.
What does all this mean? π€―
Loop follows how "Product work" is done at Google, a model many other companies also use.
Every Workspace in Loop is equivalent to a scope that is normally handled by a complete product engineering team.
It is no longer a place where users play the role of engineering lead and product lead, orchestrating AI agents.
It is a place where the user plays the role of a senior manager managing several engineering teams, which is a big shift on its own.
We are slowly moving towards autonomous AI organizations. One management layer at a time.
I expect it will still take until 2027 for labs to develop the products and models needed to make substantial progress in this area (assuming AI continues to advance in a semi-regulated format as it is now).
Monitoring πΏ
Today we're launching @arcfront with @Naman_Bhalla and deep support from @asxna.
In under a year, four of the five frontier labs became customers.
In 2025 an AI solved all 12 problems at the ICPC World Finals. The best human team solved 11. I competed in two of those finals, so I know how big a feat that is.
And yet, put that same AI inside a company as a finance manager, or in almost any other job, and it fails on day 1.
Why is that?
@rohanbhanot exactly. connection time and actual conversation time become separate budget lines. i'd want to know how quickly a disconnected session can resume before paying to keep every quiet line open.
openai shipped GPT-Live-1.1 with standby mode: $0.02/min while quiet, $0.05/min active. backend model + tools cost extra.
we've reached the part of voice agents where awkward silence has its own pricing tier.
https://t.co/cMXrHZt1CF
@0xPaulThomas@dreamDEXSomnia six personalities on the same model can still share one blind spot. the interesting test is whether their errors are actually independent, not whether they argue convincingly.
@tmaiaroto separate harness failures from agent failures in CI. a broken container shouldn't become a model score. a small set of repeatable tasks from your own workload seems more useful for regressions than another leaderboard lap.