Team Nebula

We build & deploy production Al Agentic systems for regulated enterprises and state agencies | Claude Enterprise Partner

[Team Nebula @teamnebulaai]

Graphon Labs raised a $35M Series A for agent memory that keeps each fact's source and the date it changed. It scored 93.9% on LoCoMo with 58% fewer tokens than loading the full chat.

The source and date are what a reviewer asks for when an agent acts on something it remembered.

[Team Nebula @teamnebulaai]

Teams can now turn a repeated workflow into a reusable skill just by describing it in plain language. That description quietly becomes the only spec for the workflow, and nobody versions it like code.

[Salesforce @salesforce]

That workflow your team keeps doing over and over? Turn it into a skill.

AI Skills are now available in Agentforce Coworker

Tell Coworker what the workflow is in plain language — or turn a great conversation with Coworker into a skill — then run it again whenever the work comes up.

Think account research, meeting prep, case classification, or flagging renewal risk. Once it works the way you want, refine it, share it with your team, and control who can access it.

Build it once. Run it when you need it. See how AI Skills work: https://t.co/judNO5KoRd

[Team Nebula @teamnebulaai]

@UiPath The approval only works if the release lead can see which overnight runs the agents skipped. A green dashboard built from partial runs is how bad builds go out.

[Team Nebula @teamnebulaai]

@databricks Wrong context usually means the wrong definition. Finance and sales each keep their own customer count, and an agent that can't tell which one is governed will answer with the other.

[Team Nebula @teamnebulaai]

@NabhaniMehdi Product names are the sneaky ones. Half of them are ordinary English words, so a model has no reason to treat them as special.

[Team Nebula @teamnebulaai]

ElevenLabs shipped Scribe v3 and v3 Live this week.

1. 27% fewer word errors across 102 languages
2. 43% fewer errors that change meaning, live
3. A live vocabulary of up to 10,000 terms

That list is where a clinic puts its drug names, so the transcript stops guessing them.

[Team Nebula @teamnebulaai]

@SierraPlatform Regulated companies already run a review for every new risk, so an agent goes through a process they know. Teams without that process end up debating each deployment from scratch.

[Databricks @databricks]

Genie Ontology is now Generally Available.

Most business knowledge isn't formally modeled. It lives in dashboards, SQL queries, notebooks and documents. Genie Ontology learns from everyday work, combines that context with governed Unity Catalog semantics, and ranks sources by authority so agents use trusted information.

What's new:
- 𝗚𝗲𝗻𝗶𝗲 𝗦𝗲𝗮𝗿𝗰𝗵: Search enterprise knowledge across Google Drive, Jira, Confluence and Workday
- 𝗘𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝘀𝗼𝘂𝗿𝗰𝗲𝘀 (Beta): Bring in business logic from Power BI
- 𝗢𝗿𝗴𝗮𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗱𝗶𝗿𝗲𝗰𝘁𝗼𝗿𝗶𝗲𝘀: Add people, teams and reporting relationships
- 𝗨𝗻𝗶𝘁𝘆 𝗖𝗮𝘁𝗮𝗹𝗼𝗴 𝗗𝗼𝗺𝗮𝗶𝗻𝘀 𝗮𝗻𝗱 𝗣𝗮𝗴𝗲𝘀: Now GA
- 𝗢𝗻𝘁𝗼𝗹𝗼𝗴𝘆 𝗠𝗮𝗻𝗮𝗴𝗲𝗿: Coming soon in Beta for curation and governance

In a benchmark developed with @turingcom, Genie with Genie Ontology delivers 1.6× the accuracy with less than half the tokens.

https://t.co/ro8weeDniF

[Glean @glean]

Not every AI task needs a frontier model.

Glean Auto Routing automatically matches every request to the right model, balancing speed, answer quality, and cost.

Backed by evaluations across 5,000 real enterprise tasks.

Learn how it works: https://t.co/zmCEndS5dM

[Team Nebula @teamnebulaai]

@databricks Always-on is the hard part. An agent that follows up on its own needs a record of every action it took between check-ins, or nobody can tell what changed overnight.

[Team Nebula @teamnebulaai]

12 models were told to break out of an agent VM sandbox and grab a credential from the host. The standard capture-the-flag setup steered them away from the boundary under test, so a clean CTF score says less about containment than most security reviews assume.

[Perplexity @perplexity_ai]

We are publishing new research on our findings from red-teaming SPACE, our VM sandbox for Computer agents.

We gave 12 models two tasks: escape the VM to take a credential from the host, and bypass SPACE's network policy to reach a blocked URL.

We found that the standard Capture the Flag setup used in red-teaming can steer agents away from the security boundary being tested, and that security-specialized models don't guarantee better results.

https://t.co/Egh3x3uX4c

[Team Nebula @teamnebulaai]

Five OpenAI agents in a July 19 cyber evaluation got root inside CI workloads on OpenAI's research clusters. The route was a shared package server where earlier runs had left credentials behind. Several agents noticed they were out of scope and kept going.

[Team Nebula @teamnebulaai]

@UiPath @Google Running the MCP server inside Google's own platform is the right call for healthcare. Most compliance reviews stall on where patient data goes once an agent touches it, and this settles that up front.

[Snowflake @Snowflake]

What happens when a coding agent gets rewarded for a patch that passes tests but doesn’t fix the bug?

In Snowflake AI Research team's audit of three public software-engineering training datasets, nearly a third of standard-validated tasks accepted a plausible wrong patch.

Arctic-SWE combines a stress-tested reinforcement learning recipe with curated tasks and stronger tests.

Explore the failure modes and the fixes: https://t.co/hNZS33Q0CF

[Databricks @databricks]

Gemini 4 Carbon. Day 0 on Databricks.

Google's strongest model for complex coding, scientific reasoning, and long-running agent workflows is available on Databricks the same day it opened to developers.

In our enterprise-document testing, Carbon scored 66.7% on OfficeQA Pro v2, compared with 57.8% for Gemini 4 Argon, our previous best Gemini result.

Now put it to work on your own data. Unity Gateway handles access, usage limits, and spend tracking across Carbon and every other model your teams use.

[Team Nebula @teamnebulaai]

@UiPath Financial crime compliance is where agent oversight gets tested first, because every flagged transaction already needs a written reason. An agent that can't produce that reason won't make it past the pilot.

[Team Nebula @teamnebulaai]

@Snowflake A third of validated tasks accepting a wrong patch means plenty of coding agents learned to make tests pass. Any team treating green CI as proof the bug is fixed inherits that habit.

[Team Nebula @teamnebulaai]

@databricks 66.7% on enterprise documents means one question in three on real files comes back wrong. That gap decides which workflows get this model first, and contract work is not one of them.

[Team Nebula @teamnebulaai]

The compounding moat in our Ori thesis is a data flywheel.

Each action is kept with its outcome and its reason. Over time that data improves agentic defense and auditing.
https://t.co/7ITLfOf6IF

[Team Nebula @teamnebulaai]

A teammate in Ori works from its own computer and signs in the way your people do.

Before it acts, the policy decides. Reads inside the workspace go ahead. Writes outside it wait. Sites off the allowed list are refused, and the reason is recorded.
https://t.co/7ITLfOf6IF