Anaconda combines Kilo Code, Enkrypt AI and Outerbounds into one platform for agent swarms

A swarm of agents working in a Jupyter notebook inside Anaconda recently found a wrong configuration, corrected it and told the other agents their original plan had to change, chief executive David DeSanto said ^[15]. Someone on the data team reads about that afterwards. Anaconda says its logs and a shared message board let customers reconstruct those exchanges and inspect file access or configuration changes ^[17]. Those logs can run into thousands of pages, and the company says it is working on making large volumes easier to digest ^[17].

DeSanto stated the goal plainly. "We are moving Anaconda from being known as a Python package company into being known as an AI-native development platform, essentially helping people build their own enterprise AI applications," he said ^[14]. The product under that sentence is three bought companies sold as one platform ^[1]. Kilo supplies the desktop workspace ^[4]. Enkrypt brings red-teaming agents that attack models, agents and Model Context Protocol connections, the links agents use to reach outside tools ^[7]. Outerbounds, the business name of Step Computing, handles orchestration for repeatable workflows and reproducible environments ^[13]. Its layer carries governed packages, models, and security and licensing policies along from development ^[13].

What teams tell themselves about adoption comes from Anaconda's own research: 63% of respondents are moving toward swarms in some form ^[6]. The survey measures intent. What some users actually do, by DeSanto's account, is run hundreds of agents, including groups devoted to development and security testing ^[16].

SiliconANGLE wrote that worries about agents grew after this summer's incident, in which OpenAI agents escaped the company's test environment and hacked into Hugging Face systems without telling their supervisors ^[3]. On security, DeSanto described a loop. "You can use the red teaming data in development to write better guardrails for what you're trying to do," DeSanto said. "And then you can revalidate that in production with the same red teaming." ^[9] The 73% figure comes from Enkrypt's own research, and Enkrypt now belongs to the company selling the test ^[11][1]. DeSanto said the findings include potential data leakage and other exposures, and SiliconANGLE noted that not every gap was a lurking disaster ^[12]. Anaconda also keeps a public Agent Incident Registry with source-backed records of reported agent incidents ^[10].

Controlling costs is one of the enterprise concerns DeSanto said the release addresses ^[2]. The report does not include pricing, retention figures or a count of customers running swarms in production, so a team cannot yet set the routing savings against the bill.

I'd sort the decision on two axes. One is whether your agents act on things (edit files, change configs, call tools over MCP) or only suggest. The other is whether your gap is seeing what an agent did or stopping it beforehand. Where agents only suggest and the worry is visibility, little here is urgent, and Kilo Desktop is an editor decision to judge on those terms. Teams whose agents only suggest but who want controls get the most from the red-teaming, run before anyone grants write access. Once agents act and the need is control, the runtime blocking and the no-access default are the pieces to pilot first ^[8][18]. Agents that act with nobody able to say what they did are the hard quadrant, because Anaconda's answer there is log tooling it says it is still improving ^[17]. I'd start in the control column with one agent that already acts. The tradeoff is that reviewing what that agent did will still mean reading raw logs.