@valerii_arch

none of my agent tools throw. a 503 comes back to the model as {error: true, status_code: 503} and it keeps reasoning with the failure as evidence. tool failures are just data.

@bcherny artifacts make agent output reviewable instead of buried in chat

I've been using Artifacts in Claude Code for everything: visual explanations of tricky code, system diagrams, quick previews of a few animation options, data analyses and dashboards I share with the team. They are a game changer for how I work with Claude. Can't wait to hear what you think!

@giffmana models still need a stale cache for social reputation

At least the LLMs still refer to me as a researcher rather than a shitposter🥹 That's just because they are slow to react I guess lol

@Teknium batching memory writes is where agent memory starts feeling like state

We have just merged an expanded form of the memory management tool Hermes Agent uses to save/edit/remove memories so that it can do batch operations, saving many turns of tool calls in most scenarios! Run `hermes update` in your CLI or update in your GUI to start saving now.

@salesforce simulation quality matters more than polished voice demos

@jerryjliu0 agentic search is mostly letting grep back into the building

Agentic search has moved from fixed RAG pipelines into flexible agent harnesses with access to a set of search tools: keyword search (bm25, grep regex) and semantic search. When you upload a collection of unstructured documents to LlamaParse, we expose all these tools for agents to access. Come check out our webinar on June 30th where we explore all these different tools and identify which ones work the best for agentic search: landing.llamaindex.ai/retrieval-harn…

@danshipper @every @ninklefitz biology keeps the eval loop grounded in actual experiments

we @every make investments every so often in founders building things in AI who we think are excellent @ninklefitz is one of those founders—and i am also a huge fan of tacit's mission and approach

@matthewclifford imaging speed matters less than what clinicians can trust from it

Extraordinary: full body scans in under a minute at incredible precision…

@llama_index hybrid retrieval works better once exact references matter

Vector databases or pure grep? Teams are split on the right retrieval architecture for agents.
⁣ ⁣ The reality? You need both. Semantic search for a fast first pass; grep and file reads for surgical precision when top-k chunks cut off mid-answer.
⁣ ⁣ On June 29, our Head of Engineering George He goes under the hood on the architecture decisions and dead ends behind building this harness into LlamaParse Index.⁣ ⁣ 
Register here : landing.llamaindex.ai/retrieval-harn…

@systemdesignone namespaces and cgroups explain more than memorizing docker commands

If I had to start with Docker (in 2026), I'd learn these concepts:

@IntCyberDigest credential reuse turns one breach into a self-sustaining attack loop

‼️🚨 BREAKING: 320,000 Fortinet firewall devices have been targeted in a campaign that has been dubbed 'FortiBleed'. Attackers were able to confirm 75,000 working credentials against the admin and SSL VPN interfaces. The victims include really big names like Samsung, Oracle, Spotify, Sony, and more. The data was first surfaced by researcher Volodymyr "Bob" Diachenko and analyzed by Hudson Rock and SOCRadar. The operation runs as a self-feeding loop. Attackers scan the internet for exposed Fortinet devices, then test each one against a curated list of passwords leaked from earlier Fortinet breaches and infostealer logs. Every successful login gets recorded into a verified database. They then turn each compromised box into a listening post, sniffing the traffic passing through the firewall to harvest fresh credentials, which go straight back into the scanner. The scale is large. The group ran an estimated 1.16 billion credential attempts against more than 320,000 FortiGate targets, plus 2.1 billion brute-force tries against 160,000 MSSQL servers. In the deeper intrusions they intercept SSL VPN authentication hashes, crack them on a dedicated 45-GPU cluster, and move into internal Active Directory. Diachenko confirmed full network compromises in Japan, Taiwan, Vietnam, Iraq, and Turkey, including a Turkish NATO defense contractor that had classified defense documents stolen. If you run Fortinet, act now: rotate every VPN and admin credential, enforce MFA on all external gateways, restrict management access to approved sources, segment internal networks, and audit gateway logs for unusual logins. Hudson Rock has a free domain lookup at hudsonrock.com/fortinet. Data surfaced via the Hunt Intelligence, Inc. feed.

@satyanadella long-running agents need clear approval boundaries before broader enterprise access

Copilot Cowork is now generally available worldwide, now with multi-model support! Every organization can put long-running agents to work on complex, multi-step tasks, grounded in your organization's unique knowledge and know-how. microsoft.com/en-us/microsof…

@code_star focus compounds faster when coordination cost stays near zero

Startups essentially exist when an intensity of focus is more valuable to an outcome than resources or process. It’s taken a long time for this to fully click for me. There are some ideas that really can’t be accomplished at a big tech company. It’s rarely about only talent.

@LangChain task success needs to be measured above transport-level success

An agent can return a “successful” response and still fail the task. It may call the wrong tool, skip an approval step, use the wrong context, or produce an answer that sounds right but isn’t. That’s why production agent teams need more than uptime, latency, and error rates. They need visibility into the full trajectory of what the agent actually did.

@bernhardsson sub-second sandbox startup changes how interactive agents feel

Our sandbox team has been on a crusade against every millisecond of latency and it's paying off. More cool results coming very soon!

@ericzakariasson cli for local primitives, mcp for shared authenticated integrations

mcp > cli im glad this debase is not as active anymore. i use both every day. they just serve different purposes cli for stuff the model already knows. git, gh, npm, docker, file ops. trained on man pages, and costs almost nothing in context. if im already signed in locally theres no reason to wrap it in anything mcp for most integrations. slack, notion, linear, twitter and its neat to have a protocol for all these integrations - add one server to my teams cursor and everyone gets access - auth once, persists, same locally and in cloud it also just feels better in cursor. rich icons, traceable, you can easily follow whats happening (tbh not reading that much) cli for personal, mcp for team (with oauth)

@rakyll nine days off is basically a full model generation now

Every week in AI feels like a month. My 9 days of vacation felt like 2.5 months.

@rauchg same substrate, different defaults around isolation, lifetime, and scheduling

A sandbox A function A server A build Are you getting it!? These are all expressions of the same underlying compute infrastructure. With tweaks to load balancing, concurrency, persistence, overcommit… 2026 is the year serverless and servers finally converge. With no gotchas

@IntCyberDigest unpaid security triage doesn’t scale just because the project does

❗️curl is going to stop accepting and handling security vulnerability reports, for a month. The HackerOne form and the security email both go dark July 1, reopening August 3. Maintainers cite roughly four months of relentless report pressure. Paid support contracts stay exempt. They're calling it the "summer of bliss."

@hwchase17 cost controls get useful once they connect spend back to traces