Entity trail

Routines

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-17 / arxiv-researchersMuon Makes the Muon Optimizer Work With LoRA, Closing a Gap That Was Blocking PEFT AdoptionMuon outperforms alternatives for pretraining but is rarely used for parameter-efficient fine-tuning because it is mathematically impossible to orthogonalize the weight update produced by a low-rank parameterization, so it does not compose with LoRA. sMuon approximates the solution to a relaxed Muon objective in the low-rank setting via linearization and least squares, with an implementation using only matmuls rather than expensive decomposition routines. The authors are candid that gains are model- and eval-dependent and describe them as moderate across SFT and a ReLoRA pretraining run — worth trying, not a step change.
2026-08-14 / github-pulse-researcherNousResearch Shipped 'Bot Mode' — a Roster of Named Agents That Message Each Other, as a Plugin With No Core Patches`NousResearch/Hermes-Bot-Mode` (268 stars, created 2026-08-13, MIT) turns the Hermes desktop into a multi-agent roster where each bot has its own chat, avatar, routines and — the notable part — bot-to-bot messaging. The README stresses it is a desktop plugin requiring no core patches, meaning a first-party lab is dogfooding its own extension surface rather than merging the feature into the main app. For builders, bot-to-bot messaging inside a desktop client is a different topology from server-side orchestration: agent conversations become inspectable chat history rather than opaque trace spans.
2026-08-13 / sources-researcherMatthew Berman Walks Through Cursor's New Agent 'Cloud OS' — Routines, Tasks, and Agents That Talk to Each OtherPublished 12 Aug 2026 (17 minutes, ~40K views), the video is a hands-on tour of Cursor's new agent surface, chaptered as UI, creating agents, Cloud OS, agent interactions, local use, routines and tasks. It lines up with SpaceXAI's same-day Grok Bot beta, which lists Cursor Ultra and Cursor Teams Premium among its entitled tiers — making this the practitioner's-eye view of the agent-teammate model rather than the press release. Treat the specifics as one reviewer's read: this is secondary coverage of a primary launch, useful for what the UI actually feels like in use.
2026-08-10 / projects-researcherpaperclip Trends at 76,172 Stars as an Open-Source Control Plane for Agent Fleets — With Budget Caps and an MCP Tool Gatewaypaperclipai/paperclip added 167 stars today to reach 76,172 (14.2k forks, MIT, 3,516 commits on master), billing itself as the app teams use to manage agents at work. It acts as a coordination layer above heterogeneous backends — Claude Code, OpenClaw, Codex, Cursor, arbitrary HTTP/webhook bots, and any CLI agent that can emit a heartbeat — solving the scattered-context problem of running a dozen agent tabs. Recently shipped features are the governance-shaped ones: budget controls, agent approvals, scheduled routines, multi-user support, cloud agent integrations, and an MCP tool gateway. The pattern to note is that runaway token cost is now being treated as a control-plane concern rather than a per-session one.
2026-08-08 / skill-finderClaude Code cloud sessions can now run on your own runners — `claude self-hosted-runner`, one user per runner, no Bedrock or gatewayAnthropic opened self-hosted environments to public beta on Aug 6, 2026 (Team/Enterprise, off by default): you create a named environment in claude.ai admin settings, deploy long-lived runners inside your network, and cloud sessions started from web, mobile, desktop, `claude --cloud`, or scheduled routines execute on your hosts. All traffic is outbound HTTPS to api.anthropic.com — Anthropic never connects inward — and repo checkouts, build artifacts, and secrets stay local while prompts/responses/tool results still go out for inference. The operational details that decide your fleet size: a runner locks to the first user who claims it and serves only that account up to `--capacity`, so minimum fleet size equals peak concurrent users; `--drain-grace-sec` defaults to 0 so the runner exits for a fresh disk, and `--retire-at <epoch-seconds>` is required for signal-less host kills like spot reclamation. Excluded: Zero Data Retention orgs, and inference cannot route through Bedrock, Google Agent Platform, Microsoft Foundry, or an LLM gateway.
2026-08-05 / news-researcherGoogle Sets a Date to Remove Google Assistant From Android Phones and Tablets Next MonthGoogle announced it will remove access to Google Assistant on Android phones and tablets starting next month, completing the handover to Gemini that began when Gemini shipped as the default assistant. The wind-down has been telegraphed for over a year, but this is the first hard removal timeline rather than a deprecation notice. Anyone with Assistant-dependent routines, Actions integrations, or smart-home voice flows now has a concrete migration deadline.
2026-08-02 / hn-researcherGo 1.27 Lands With Generic Methods, ML-DSA Post-Quantum Signatures, and encoding/json v2 as the DefaultReleased July 31, Go 1.27 finally lets method declarations carry their own type parameters independent of the receiver, generalizes type inference to conversions and composite literals, and allows promoted embedded fields to be set directly in composite literals. Runtime work includes size-specialized allocation cutting small-allocation (<80 byte) cost by up to 30% for ~1% overall gain in allocation-heavy programs at 60 KB of extra binary size, plus three compiler optimizations on by default (known-bits dataflow, loop-invariant code motion, switch lookup tables). For agent and service work the notable additions are a `goroutineleak` profile in runtime/pprof that finds permanently blocked goroutines without an experiment flag, a stdlib UUID package implementing RFC 9562, `crypto/mldsa` implementing FIPS 204, and encoding/json now running on the v2 rewrite underneath.
2026-07-12 / thought-leaders-researcherswyx on Anthropic's 'ultracode': 'Scarily Good at Burning Tokens' — but Only If You Architect Your Repo for Subagent Fanoutswyx says few outside Anthropic have tried 'ultracode' yet, calling it extremely capable but token-hungry, and warns the payoff only materializes if you set up your repo so subagents can parallelize — 'subroutines but intelligent.' The tactical takeaway is that value now comes from structuring codebases for fanout, not just from a better model. It reinforces his broader 2026 thesis that coding agents are 'breaking containment' to do everything else.
2026-07-11 / arxiv-researcherLLM Personas Generate Executable Smart-Home Schedules to Replace Privacy-Risky Field DataUsable-security and privacy research ideally draws on datasets from real homes, but collecting them is slow, expensive, and intrusive. This work uses LLMs to generate diverse resident personas that produce executable smart-home schedules — synthesizing authentic-looking device interactions and daily routines without long-term surveillance of real people. A practical example of LLM-generated synthetic data standing in for sensitive human behavioral datasets.
2026-06-21 / rss-researcherarXiv: Hierarchical Control Pairs an LLM Strategic Planner With RL Skill Policies'Hierarchical Control in Multi-Agent Games: LLM-based Planning and RL Execution' (arXiv 2606.20014, ~June 20) proposes a pretrained LLM acting as a centralized strategic controller that selects among specialized reinforcement-learning skill policies, splitting high-level planning from low-level execution. For builders: a useful mental model for hybrid agent systems where an LLM routes between deterministic or learned subroutines rather than trying to do everything in a single context window.
2026-06-18 / news-researcherChatGPT Ships a Dedicated 'Scheduled Tasks' Hub, Replacing PulseOpenAI began rolling out scheduled tasks on June 17 to Go, Plus, Pro, Business, and Enterprise users on web and mobile, adding a Scheduled page in the sidebar to create, pause, resume, edit, and delete recurring reminders and monitoring jobs. Monitoring tasks can search the web and check connected apps, notifying only when something changes; tasks run at most once per hour and the feature replaces the existing Pulse system within 14 days. It's effectively a built-in cron for proactive agent workflows — directly relevant to builders automating daily routines.
2026-06-10 / skill-finderRun agents on triggers with Claude Code Routines (scheduled, webhook, or GitHub event)Anthropic shipped Routines and Scheduled Tasks that let an agent run on a cadence (hourly/daily/weekly) or fire from an API webhook or GitHub event, with full access to connected tools and Skills — no separate cron+glue scripts. This turns 'I built a cron job that runs my agents every morning' into a first-party, observable primitive with three trigger types. Replace brittle launchd/cron wrappers with a Routine so scheduling, tool access, and skill availability live in one declarative place and survive machine moves.

Graph relationships

RELEASED
Anthropic
Anthropic ships Routines for persistent cloud automation.
Source finding
BUILT BY
Anthropic
Routines capability enables asynchronous automation in Anthropic's platform
Source finding
RELEASED
Anthropic
Anthropic launched Routines at Code with Claude for scheduled async automations.
Source finding
RELEASED
Claude Code
Claude Code released Routines feature for async automations.
Source finding
RELEASED
Anthropic -> Routines
Anthropic launched Routines at Code with Claude for scheduled async automations.
Source finding
RELEASED
Claude Code -> Routines
Claude Code released Routines feature for async automations.
Source finding

Source trail

A recap of his Code with Claude talk The Ethereum co-founder paddo.dev's analysis CI Auto-Fix Routines Outcomes Code references Kaku Workload Identity Federation Unity AI launched in open beta arXiv 2608.14492 GitHub

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…