Micheal Benedict (@micheal) header
Micheal Benedict (@micheal) profile picture

co-founder @stacktrace_ai (YC F26) | Real-time steering of AI agents Previously eng @ databricks, airtable, pinterest & twitter.

California, USA
Joined September 2008
Photo by Micheal Benedict (@micheal)Photo by Micheal Benedict (@micheal)Photo by Micheal Benedict (@micheal)Photo by Micheal Benedict (@micheal)Photo by Micheal Benedict (@micheal)Photo by Micheal Benedict (@micheal)Photo by Micheal Benedict (@micheal)Photo by Micheal Benedict (@micheal)Photo by Micheal Benedict (@micheal)
📣 Quick life update: Very excited to have started something w/ @vinod and Prashanth -- Introducing Stacktrace (@stacktrace_ai)! Some context: At Airtable & Databricks, I worked on helping teams move faster without compromising reliability or security. With unattended agents (especially at scale), the problem compounds: they drift, get stuck or take actions you did not intend. At scale, teams cannot keep watching the work they thought they had delegated. I want to give agents more responsibility. However, to do this effectively -- we collectively need better systems & mechanisms that instill confidence on agents. This is exactly why we are building Stacktrace, a platform that can. 1/ See & explain what agents do 2/ Control what they allowed to do 3/ Intervene when they cross a boundary or drift from the task @vinod, Prashanth & I have spent much of our careers building and operating infrastructure & security systems at scale. We want to help define "what good looks like" for agentic infra: telemetry that explains what is happening, controls to steer execution and the ability to intervene and manage drift. We believe these foundations are critical to running unsupervised intelligence with confidence, and entrust them with consequential work. We are working with a few trusted design partners running agents at scale, locally or unattended in the cloud. If you are interested to learn more or try out Stacktrace within your organization, DM me. -mb
2
14
1,670
If someone says they offer, ZDR it is always going to come with * ask them how is it handled in each part of the stack. Ex, everything from ingress, to how they handle data in transit, ephemeral storage layers, online and offline data sinks, operational and security logging pipelines, use with sub processors, etc. IMO ZDR is a dance between GRC and Eng where time-to-market forcing sub optimal/incomplete implementations protected by nuanced language on the contract.
We seeded 2m canaries into popular AI services and tracked where they were stolen and abused. This identified: - Three compromised providers - Resellers stealing prompts via built-in logging - Providers ignoring ZDR & privacy flags dragos.com/blog/ai-supply-ch…
74
this is pretty wild, if true we have reached a point where agent output is rarely monitored by humans. turning off safety in models will open up a new class of cyber attacks (self-inflicting types) that we simply don't have any systems to actively monitor & deter in real-time
Looks like @PrimeIntellect was compromised in a very strange way. No public comms from them yet, but disclosed by OpenRouter to affected folks (such as myself)
82
the og "sandbox" company is making a comeback
A new Docker, but a familiar story. Today we're introducing a new look, a new focus, and a refreshed Moby. Same Docker you've trusted for over a decade, now built to run on trust in the agentic era. We can't wait to show you. Read the full story: bit.ly/4iX61ek
1
78
Our competitor launched a product with the tag line "steer on every agent action, on every platform, in real time." flattered
63
Still waiting to hear which enterprise (that has real stake) has switched purely because of price. Most of the new sandbox companies seem to be PLG. I have heard capacity being a problem/bottleneck when wanting to run throwaway/batch workloads -- but again, given how cheap it is to build and tailor sandbox infra to your use-case, i expect strong consolidation here.
Sandboxes are coming now to your own infrastructure. The age of Daytona, Modal and E2B dominance on sandboxes is very close to being over, for three main reasons: * Market commoditization * Cheaper sandboxing options: Agent37.com and boat.dev as prime examples * Community interest pushing towards OSS solutions easy to run on-premise My bet: many of the "big" sandbox companies will become Zombie startups in no time
1
105
I am not an Airbyte user, but have been on the other side of such incidents -- reminder that these are extremely stressful events for everyone involved coordinating with expert threat & detection teams, possibly law enforcement, understanding entrypoint & extent of breach, patching the gaps (be it in application code, revoking keys, ruling out 3rd party svcs), etc.
Airbyte has identified a security incident that we believe is part of a broader, global cyberattack. Fewer than 1% of our customers were affected, and we have already contacted each of them directly. Our security team detected the attack, blocked it, and took Airbyte Cloud offline, preventing the attacker from maintaining access to our systems. Our investigation is ongoing. Because protecting our customers’ data is our top priority, Airbyte Cloud will remain offline until we are confident it is safe to restore services. We will share further updates as we learn more. Customers with questions can contact [email protected].
2
445
I love the iPhoneDuo. It is quite the versatile device with a lot of potential. Ended up building custom apps/porting games for myself. Here is an example of a game called Claw (by Monolith Games, circa 1997) running on the Duo. It looks gorgeous!
1
182
We built Stackbox for our own CI and benchmarks at @stacktrace_ai . Now opening it up for others to try. Stackbox: OpenRouter for Sandboxes One API across sandbox providers. Route for cost or faster starts, and seamless handle provider capacity and availability. Running agents or evals? Give it a spin: app.box.stacktrace.ai
2
103
hugops to the entire team dealing with this incident If you have been in these type of rooms, you know the immense pressure and stress associated
Replying to @MSFT365Status
We’re seeing positive results for service stability and increased capacity as a result of our latest traffic optimization actions. We're implementing these actions more broadly across affected regions as we seek to attain persistent recovery. A more comprehensive update can be found at MO1550821 in the Microsoft 365 admin center, as well as status.cloud.microsoft if the admin center is not available to you.
159
We have always used @railway since we started @stacktrace_ai and continue to build our business on it. It is one of the very few companies with strong principles, who strives to do the right thing and lives up to that expectation.
As mentioned in our post mortem earlier this week. All losses from stolen keys will be made whole. We wanna thank our partners for helping make this happen. An update on the above, and how we’re doing it, below.
3
259
Anthropic's product strategy is very reminscient of what Databricks did (but a different scale) The core product was compute ($/dbus). While the business started with ETL, DBSQL really made compute accessible to all. Products like Serverless-* to Genie, all build atop of this foundation while products like VectorDB, postgresql (neon), model serving & fine tuning offered newer SKUs while adding to compute. Anthropic seems to be on a similar trajectory with Docs, Sheets, Data and now Meetings. Each one brings a new kind of user who add to tok usage. Couple this with ancillary svcs like managed agents, artifacts, sandboxes that continues to add to the core offering (i.e toks)
Introducing Claude Meetings: add Claude to your team’s video call and it builds docs, whiteboards, and other artifacts with you while you’re still talking. Starting to roll out in beta on Team and Enterprise plans.
146
Congrats on the launch! Wish there was a solution earlier 😓 @vinodkone built a pretty reliable sandbox router for our internal use (CI and benchmark infra). This worked remarkably well that we just opened it for free. app.box.stacktrace.ai/ * Works with external (and your setup) - daytona, e2b, modal, blaxel, gke, aws, etc * One CLI + API to manage em all. Easy add any provider there as long as they talk open sandbox * Configure placement strategy (cost, capacity or "i am feeling lucky") * Logs + Audit trail
Getting your first sandbox is easy. Getting thousands of them isn't. On Thursday I'm showing what we built to fix that: one API for every provider, with routing and failover built in. Thursday, 10AM PT luma.com/euvgmp9h
143
if you are solely on cloud, this is a no-brainer If you use claude, codex & other llms -- then, this does get complex as there is no "agent-native cloud stack" to deploy
Self-hosted Claude Managed Agents is now in private beta. You can host the agent harness, sandboxes, and session history, in your own cloud or on-premises, and choose where requests to Claude are sent.
94
talk about timing
CEO of the Year has a nice ring to it. Congratulations to AMD Chair and CEO @LisaSu on this well-deserved recognition from @YahooFinance! 👏 yhoo.it/4kdfEYI
89
sandbox infra ain't cheap have been testing out gcp & aws's offerings -- gcp outright costed us over $500+ to simply setup a 200vCPU cluster
82
We gave claude & codex a few rules in plain English. How confident are you that they'll follow them? → Credential Egress: Never send a credential off this machine → PII Data: Never send a customer's personal data to a third party → DLP: Don't put company data on the internet without a login In our early benchmark, we tested* 3 rules against 9 frontier models: Fable 5.2, Opus 5.5, Sonnet 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, GLM-5.3, Kimi K3.1 Pro and DeepSeek V4.1 Flash. Spoiler: → Without any guardrails, agents violated 60% (17 of 27) sessions → With Claude and Codex's default auto permissions mode, agents violated 30% (8 of 27) sessions → With Stacktrace, there were zero violations regardless of the model & harness detailed write up coming soon * every one is a separate session. Agents don't take the same path
7
1
312
one more here: reddit.com/r/Notion/comments… all the more reiterates that system of record type services *should* own your inference stack. The only nuance here is is not all data is equal in value (not to mention who gets access to it and what can be done with it); however, every form of data leak erodes customer trust far more than you think given cost of building is zero, don't be surprised if your customers move to something built in-house (where again, they own the system of record)
The Fireworks hack exposed Notion agent conversations :(
2
1
5
3,333
sounds like fireworks internal infra was compromised; however, no details on the how .. that said, the fact that they went from a dev machine to a production accounts is quite jarring
128