Reading List – 2026-11-20

Agent permissions reach beyond the approval prompt. Policy changes can grant new access, a room invitation can expose old messages, and a deferred decision depends on resuming the right session. I find that surrounding machinery as interesting as the agent itself. Separately, PostgreSQL upgrades can change search behavior in ways that call for index and data maintenance.

OpenShell

OpenShell treats changes to an agent’s access policy as something to verify, not just approve. It uses formal verification to flag newly granted access, such as reaching a new host with credentials, for human review. I’m interested in that combination of runtime enforcement and review: the sandbox constrains current behavior, while the policy check makes changes to those constraints visible. Teams still need to decide whether the access being requested is appropriate.

GitHub – ai-openshell/openshell: OpenShell is the safe, private runtime for autonomous AI agents.

Token fraud is here and it’s BAD (YouTube)

Free trials for AI agents can become a channel for reselling or commercially exploiting someone else’s token budget. Claire Vo reports users cycling through trials and promo codes to route requests through ChatPRD, including for a Roblox and Fortnite map studio. What catches my attention is the scale of the incentive: free usage can support another business. Vo is winding down free trials, trading an easy evaluation path for relief from the abuse.

Token fraud is here and it’s BAD (YouTube)

Building always-on agents

A deferred tool call lets an unattended agent ask for approval without tying up its worker. The run ends mid-call, and the user’s reply returns through the inbox to resume the saved session. I like the separation: other events can run while the decision waits, rather than putting the whole agent on hold. That makes session persistence and correct resumption part of the permission system, not just a convenience for managing context.

Building always-on agents

Introducing Tool Retrieval in the Gemini API

What catches my attention is that Gemini keeps tool names and descriptions visible while deferring their parameter schemas. The model requests a tool by name, and the API loads its schema within the same request. Google says loaded schemas are billed as input for the rest of the conversation, so savings depend on which tools get used. For a large tool library, measure against your own traffic and keep descriptions clear enough to support selection.

Introducing Tool Retrieval in the Gemini API

PostgreSQL 19 release and upgrade behavior

A PostgreSQL 19 upgrade can change full-text search and pg_trgm results when a database’s locale provider classifies characters differently from its LC_CTYPE setting. These functions now use the database’s locale provider rather than always using libc. I’d pay particular attention to databases using ICU or the builtin provider with LC_CTYPE set to C. For affected databases upgraded with pg_upgrade, the release notes recommend reindexing full-text and pg_trgm indexes and recomputing stored tsvector values.

Disabling a new PostgreSQL optimization doesn’t turn a version 19 server into version 18. The tour uses older SQL and configuration switches to illustrate differences in behavior, but recommends actual 18 and 19 servers for timing comparisons. I like that distinction: matching data, settings, cache conditions, and hardware matters when you’re deciding whether an upgrade improves your workload.

PostgreSQL 19 · PostgreSQL 19: A hands-on tour with runnable browser examples

Llms.txt and MCP didn’t move the needle

Version-matched instructions in an agent’s context can correct patterns it learned from older package releases. The Backpack maintainer reports that their Laravel Boost plugin helped more than a docs-only MCP server, which agents often didn’t call. I like this suggestion for making packages more agent-friendly: put current guidance where the work starts, then organize it by topic so agents can load the details they need.

Llms.txt and MCP didn’t move the needle

Pepper in group chats and agent rooms

A shared chat gives people and agents the same context without passing answers between separate conversations. Pepper says an invited agent can read the entire room backlog, including messages from before it joined, but not rooms it hasn’t joined. That history access catches my attention: it makes the agent useful immediately, while making the invitation an access decision for past conversations as well as future ones.

A group-text assistant can join people where they already coordinate, but staying in the existing chat doesn’t eliminate onboarding. Pepper replies to the whole MMS thread, yet each person it answers needs an account; unrecognized numbers get a signup pointer instead. I find the room-size behavior more interesting: beyond eight people, Pepper stays quiet rather than answering only part of the group. That makes the limit important to surface when a group grows.

Pepper lets agents with a shell join rooms as full participants, each with its own identity. What catches my attention is that the same room can include both coding agents and monitoring scripts. Separate identities could make participants easier to distinguish when maintaining those automations, though the source doesn’t explain how permissions or attribution work.

How to collaborate with people and AI agents in the same chat room · How to add an AI assistant to a group text · Anything With a Shell Can Join the Room

Inkling 2 Mini

Looping gives a model more computation between tokens, but also moves more reasoning outside the text a monitor can inspect. Inkling’s researchers found that a variable-loop model produced shorter chains of thought that remained legible, with no clear increase in successful stealth attacks as loops were added. What catches my attention is that a weaker monitor still missed significant attacks when it could see only tool calls and outputs. For coding agents, the monitoring setup matters alongside the model architecture.

Inkling 2 Mini

Alibaba Open Code Review

Alibaba’s Open Code Review puts file selection and bundling in deterministic code, while leaving context gathering and review judgments to an agent. What catches my attention is the division: related files share a review unit, but each unit gets an isolated context. The video reviewer reported a 21-minute run using about a million tokens, then a one-minute review on a smaller change with lower effort. Review depth and cost look worth tuning for routine CI versus more thorough checks.

alibaba/open-code-review · This Open Source Code Reviewer Is Actually BRILLIANT

Language Models Can Rewrite Their Own Weights

A model can apparently improve its own accuracy by editing its weights without gradients. What catches my attention is how the 27B model compared hidden states for completed answers and answers that ran out of tokens, then weakened the direction separating them. The authors report a 6.6-point gain on 490 unseen problems, with no added cost per answer. The result is limited to one model family, and smaller models failed to improve themselves.

Language Models Can Rewrite Their Own Weights

Leave a Reply