There's a lot of pretending in the AI industry right now. Products claim AI does more than it does. Benchmarks get gamed. "AI-powered" gets slapped on features that are a regex and a GPT call. The demos work; the products sometimes don't.
(Yes, I see the irony of writing this from inside a company building an AI product.)
An honest capability list is more useful than a hype list, so here's ours.
What agents do well
Right now, AI coding agents are excellent at writing boilerplate, translating a clear spec into working code, finding bugs in code you describe to them, suggesting refactors you then review, and writing tests for code you give them.
They're genuinely bad at knowing whether they're solving the right problem. At catching ambiguities you didn't notice yourself. At organizational context, the "don't touch that, it's a deliberate workaround" kind of knowledge. And at working reliably without a human who knows what done looks like.
This matters because a workflow that expects agents to do the second list fails in ways that are hard to predict and expensive to recover from.
The review gate exists because agents genuinely need humans, and we'd rather build that in than pretend otherwise.
Model-agnostic, agent-agnostic
We don't know which model will be best in six months. We don't think anyone does.
So hilos stays neutral. Connect Claude Code, Cursor, Antigravity, or your own scripts over MCP. Hosted agents run Claude by default, because right now we think it's the best option, but that sits behind a model router and can change without rebuilding anything.
Some of the companies betting everything on one model will be right. The ones who are wrong will be wrong very fast. Neutrality is the one position a model-owning incumbent can't take, so we're taking it.
The transparency commitment
Every agent action in hilos is visible to the team.
When an agent opens a PR, you see what it was asked to do, what it decided, where it got stuck, and what it skipped. The report card comes from the agent itself, generated while it works. You can disagree with it, redirect it, reject it. The conversation stays, so months later you can still see which decisions were made and why.
What we don't do
Nothing merges automatically. And we don't train on your team's conversations... the policy is written in plain language on the privacy page.
These aren't toggles waiting for an opt-in. They're not on the roadmap at all, because they'd undermine the thing that makes hilos worth using: people direct, people approve.
Where this goes
The goal is teams that trust their agents enough to hand over more autonomy over time, because the trust got built incrementally, in a workspace that kept everyone in the loop.
The honesty has a practical side too. When an agent does something unexpected, and it will, you want a team that trusts the system enough to investigate instead of panic.
We'd rather build that honestly than pretend we're further along than we are.