Josh Devon

@Al_Grigor We open sourced a policy engine that uses policy-as-code to create deterministic hard boundaries that stop the agent from making mistakes like this, even while running in YOLO mode: github.com/sondera-ai/son… And you can read the write-up of this approach here: securetrajectories.substack.com/p/hooking-codi…

@jonsakoda I learn more in 30 seconds hearing @RGB_Lights talk than even @claudeai.

2026 is only 80 days old and cyber warfare has become the biggest news story in tech The US military deployed a secret “discombobulator” weapon to black out Caracas and neutralize Venezuela’s defense 👀 Iran-linked hackers retaliated against military airstrikes by wiping 200,000 devices at one of the US’s largest medical equipment companies The US is turning the dial to “11” and pushing new limits - what’s myth vs. reality and where do we go from here? We are excited to host our annual “Proudly Offensive” meetup at RSA next week, a live podcast and unfiltered conversation about the changing US cyber defense posture with our friends who know it best: Pat Gray, founder of Risky Business Andy Boyd, Former Director, CIA Center for Cyber Intelligence Rob Joyce (@RGB_Lights), Former Director of Cybersecurity, National Security Agency Please join our community of founders and early adopters at the Decibel "Founder Festival" — we're excited to see you all soon 🌴 Check out our full lineup and register at the link below

We’re launching LangSmith Fleet today! There are some primitives in Fleet that I think will be very useful in a future where agents do a lot of the world’s work - Agent Identity: as more work is specified by humans but done by agents, we need identity + security models that reflect that. This means attaching credentials and connections specifically to agents that control what they can read/write and what surface areas they interact with (github, slack, etc) - Agent Self-Improvement: There’s more agents that we’ll use over week, month, year time-scales. Memory and the ability for agents to edit themselves means agents mold to your tasks over use. Because they’re hooked up to LangSmith, this also lets your team generate evals to ground your changes in metrics over time. - Integrate with the World: It’s kinda funny how hard (edge cases!!) it is to connect agents to external tools like Slack and GitHub. Part of it is auth but part is good context engineering to help agents actually use the connections to do useful work. Context engineering around external systems is tricky, Fleet does a lot to help with that process! - Sharing Agents: Memory + Shared surfaces of access means that agents improve over time with your team but also a useful agent built by one team is immediately helpful to other orgs! Coding Agents are work engines and helpful to marketing as much as engineering - Harness Engineering Agents need to adapt their tooling, prompts, memory to the task at hand. All of this is adapting the agent harness for the work. Now agents help you help them. Fleet is a tool for using agents but also building the harness around agents to make them great at the work you want to do. A lot of products exist by making this loop tailored to specific tasks Excited to see people try this, more agents to build and more work to do 🫡

@ProulxKerem @runpensar Awesome @ProulxKerem and @kylebhiro !! Great to see these capabilities, absolutely critical to get ahead of attackers with this tech.

Our autonomous pentesting agent just outperformed the two most popular open source offensive security agents on a benchmark of 60 modern, defense-enabled web apps. Battle-tested in production against our customers' environments from startups to financial institutions, Apex consistently finds and exploits critical vulnerabilities other agents and humans miss. Today we're releasing it open source alongside our internal benchmarks.

Founders, let’s get real…. The old playbooks don’t work anymore. You can’t follow conventional wisdom. Your advisors haven't seen this before The path ahead is unpaved and nobody knows the answers but you 👀 Sometimes you need inspiration to make hard decisions in your startup, and I often get mine from Bipul Sinha (@bipulsinha), founder and CEO of Rubrik We'll sit down with Bipul at RSA next week to talk about how AI is changing everything for founders, his journey from zero to IPO, and his formula for finding product-market fit over and over again 🔥 Looking for new ideas and fresh vibes? Please join our community of founders and early adopters at the Decibel "Founder Festival" — we're excited to see you all soon 🌴 Check out our full lineup and register at the link below:

@jonsakoda Looking forward to it @jonsakoda and @bipulsinha!

What separates an AI agent enterprises actually trust from one that quietly breaks everything? Read @joshdevonai's contribution in the Winter 2026 issue to find out. Flip through to read some excerpts, and visit 👇 issuu.com/aicybermagazin… to read the Winter 2026 issue.

@joshdevonai What separates an AI agent enterprises actually trust from one that quietly breaks everything? Read @joshdevonai's contribution in the Winter 2026 issue to find out. Flip through to read some excerpts, and visit 👇 issuu.com/aicybermagazin… to read the Winter 2026 issue.

@joshdevonai What separates an AI agent enterprises actually trust from one that quietly breaks everything? Read @joshdevonai's contribution in the Winter 2026 issue to find out. Flip through to read some excerpts, and visit 👇 issuu.com/aicybermagazin… to read the Winter 2026 issue.

@joshdevonai What separates an AI agent enterprises actually trust from one that quietly breaks everything? Read @joshdevonai's contribution in the Winter 2026 issue to find out. Flip through to read some excerpts, and visit 👇 issuu.com/aicybermagazin… to read the Winter 2026 issue.

What separates an AI agent enterprises actually trust from one that quietly breaks everything? Read @joshdevonai’s insightful contribution in the Winter 2026 issue of @aicybermagazine to find out. Flip through to read some excerpts from his piece and visit 👇 issuu.com/aicybermagazin… to read the Winter 2026 issue of @aicybermagazine.

What separates an AI agent enterprises actually trust from one that quietly breaks everything? Read @joshdevonai’s insightful contribution in the Winter 2026 issue of @aicybermagazine to find out. Flip through to read some excerpts from his piece and visit 👇 issuu.com/aicybermagazin… to read the Winter 2026 issue of @aicybermagazine.

@chrisdietr @brian_scanlan Our harness runs at the trajectory layer, evaluating the agent’s behavior pre- and post-turn, you can read more about how it works in our docs here: docs.sondera.ai/concepts/stage…

@brian_scanlan @joshdevonai agree on that one. which layers do you interact with? i have arrives also to use my shell config to capture commands that the agent should not run to capture on that level. completely agent independent.

We've been building an internal Claude Code plugin system at Intercom with 13 plugins, 100+ skills, and hooks that turn Claude into a full-stack engineering platform. Lots done, more to do. Here's a thread of some highlights.

What separates an AI agent enterprises actually trust from one that quietly breaks everything? Read @joshdevonai’s insightful contribution in the Winter 2026 issue of @aicybermagazine to find out. Flip through to read some excerpts from his piece and visit 👇 issuu.com/aicybermagazin… to read the Winter 2026 issue of @aicybermagazine.

What separates an AI agent enterprises actually trust from one that quietly breaks everything? Read @joshdevonai’s insightful contribution in the Winter 2026 issue of @aicybermagazine to find out. Flip through to read some excerpts from his piece and visit 👇 issuu.com/aicybermagazin… to read the Winter 2026 issue of @aicybermagazine.

@AlexReibman Biggest concern was risk and security. If you sandbox, it’s not useful, and if not, it’s lethal trifecta waiting to happen. We open sourced a research extension outlining some of the risks and mitigating them here using policy as code: blog.sondera.ai/p/openclaw-rm-…

Question to everyone who tried OpenClaw but hated it: What didn't you like about it?

@brian_scanlan Yeah for sure would love feedback!

@joshdevonai That looks very good. Convincing agents to always follow instructions can feel pretty futile. I'll give it a go!

@brian_scanlan Also can see the write-up we did here on this talk we gave at Unprompted describing the harness and policy engine: blog.sondera.ai/p/hooking-codi…

@brian_scanlan Super cool @brian_scanlan. If at all helpful, we open sourced our agent harness and policy engine to create deterministic controls at the action-layer using policy as code in Claude Code and the other coding agents: github.com/sondera-ai/son…