Agents on laptops are easy. Agents running safely on your real codebase, all day, are not.
Every company will run a software factory. Today we shipped the one built for enterprises ๐ค๐
#AIAgents#SoftwareFactory
Today weโre launching Islo Enterprise Software Factory๐
Coding agents can generate code. Safely automating engineering work across complex enterprise codebases requires much more:
Real environments: Agents work in persistent, isolated environments with the toolchains, dependencies, services, and compute needed to build and run the actual software.
Deterministic controls: Probabilistic agents operate inside explicit factory lines. Deterministic stages run builds, tests, and policy checks. Failures route work back for another attempt, while risky actions pause for human approval.
Evals: Teams can test changes to models, harnesses, and prompts against their own historical engineering work before rolling them out โ measuring quality, side effects, attempts, and cost in real environments.
Thatโs the difference between running an agent and operating a software factory.
Teams can use Islo to turn work like refactoring, QA, bug fixing, and feature delivery into continuous, verifiable factory lines without locking themselves into one model or harness.
Weโre also announcing $15M in new funding led by @insightpartners and a deep infrastructure partnership with @databricks to power inference for our enterprise customers.
Looking forward to seeing what you build! https://t.co/ovwb02TsNo
Running 42 FFmpeg tools locally with no API key sounds clean, until you remember the skill still shares a disk with the rest of the machine. A missing cloud key doesn't make the rest of the secrets safe. I'd treat a video skill like any other MCP, and give it a tight box rather than the developer's laptop.
A board for ten agents is useful, right up until two of them share a checkout or a set of credentials. Separate worktrees help with the files. They don't help if the processes still share a machine. At @islolabs each run gets its own hardware-isolated microVM and is snapshotted, so a parallel job can be rewound without taking the others with it.
A lawsuit over agents leaving their sandbox is a sign the isolation story has moved from research talks into court. Container boundaries were a poor fit for an agent with a shell and a goal. At @islolabs each run sits in a hardware-isolated microVM and is snapshotted, so a breakout is harder and a bad run can be rolled back.
The useful line is that builds and tests decide by exit code, not by an agent saying the work is done. Self-hosting and keeping the keys on your own infra is the right instinct too. I'd still want each of those roles on a throwaway machine, because a self-hosted loop can still share one environment across plan, code, and test. At @islolabs that machine is a hardware-isolated microVM you can stand up on any infrastructure, and the run is snapshotted if a loop goes wrong.
If attackers can already turn a patch into an exploit with no one in the loop, the same class of agent is sitting inside a lot of companies with repo access and leftover keys. The patch race is a defender problem. The coding agent on a developer's machine is a different kind of risk, and it should run without the credentials that would make a hijack worth it.
Yes, and the part I'd finish first is the runtime underneath, before any of the ticket routing. A factory only holds together if a run can't touch what it wasn't given, and if you can put a bad one back. A lot of setups still fail that test the moment you kill a job halfway through.
The bitmap itself is a small leak. The bigger issue is the assumption underneath it, that the coding agent lives on the same OS as the developer. Once that is true, the machine is already a shared place, and a diagnostics pipeline is just one more reader. I'd rather treat the laptop as a client and put the run on infrastructure the team actually controls.
Folder-level permissions on the Mac are a start, but they still assume the agent belongs on the laptop in the first place. An agent that can read the repo can usually find a .env if it shares a disk with one. At @islolabs each run gets its own hardware-isolated microVM, and the real credentials stay outside the agent, so Full Disk Access on a developer's Mac never has to be part of the setup.
Yes, and credentials are already proof the pattern works. At @islolabs the real keys stay outside the VM and only get used at the boundary, so the agent never holds them. Decisions could follow the same path. The run checks before a write or commit instead of carrying a copy that goes stale.
The plumbing is the easy part. The hard part is authority. Someone has to say which decisions are binding and which were just one engineer's preference in a Claude Code session. Without that, the service drifts the same way the markdown does, only somewhere harder to see.
Are you seeing teams actually try this yet, or is everyone still at the markdown stage?
The useful part is that the agent can discover a tested step, fill in the inputs, and validate the config before it ever pushes. That's a lot safer than letting it invent a setup script and finding out in the pipeline.
I like that setup-aws and setup-docker are just functions the agent has to configure, not shell it gets to improvise.
@rauchg If the next users are agents, the product has to be something an agent can operate without a human sitting on every step. Evals are how you find out it can't. The Vercel write-up on testing that loop before you ship is the part I'd steal.
The 18 malicious skills will get all the attention, but Snyk's haystack line is the scarier one. Most of these skills aren't written to attack you. They just happily reach for wallets, files, or hardcoded credentials.
At @islolabs the real credentials stay outside the agent, so a skill can't leak a key the run never had. That doesn't make a bad skill safe, but it does mean one install is less likely to take the rest of the system with it.
@ThePrimeagen Point 2 is the one that sticks. If we put less work into the PR than we expect a reviewer to put into reading it, that's just a tax on the next person.
@CloudNativeFdn OpenShell putting isolated sandboxes under live scrutiny is the right problem to work on in the open.
At @islolabs each agent run gets its own hardware-isolated microVM, and the real credentials stay outside the agent. Curious what blockers they hit on Thursday.