There's a lot of pretending in the AI industry right now.
Products claim AI is doing more than it is. Benchmarks are gamed. "AI-powered" is slapped on features that are a regex and a GPT call. The demos work; the products sometimes don't.
(And yes, we're aware of the irony of writing this from a company building an AI product.)
But we think honesty is a competitive position, not just a value.
What AI agents actually do well — and don't
AI coding agents, right now, are excellent at:
- Writing boilerplate and repetitive code
- Translating a clear spec into working code
- Finding bugs in code you describe to them
- Suggesting refactors you then review
- Writing tests for code you give them
They're genuinely not great at:
- Knowing whether they're solving the right problem
- Catching ambiguities you didn't notice yourself
- Understanding organizational context ("don't change that, it's a deliberate workaround")
- Working reliably without a human who knows what "done" looks like
This matters. If you build a workflow that requires agents to do the second list, it will fail in ways that are hard to predict and expensive to recover from.
We built the review gate because we believe agents need humans. Not as theater. As genuine oversight.
Model-agnostic, agent-agnostic
We don't know which AI model will be best in six months. We don't think anyone does.
So hilos is model-agnostic and agent-agnostic. You can connect Claude Code, Cursor, Windsurf, or your own scripts via the MCP server. The hosted agents default to Claude right now — because we think it's the best option — but that can change without rebuilding anything.
Companies betting hard on one model, one workflow, one paradigm... some of them will be right. But the ones who are wrong will be very wrong, very fast. We'd rather hold a position the incumbents structurally can't: genuine neutrality.
The transparency commitment
In hilos, every agent action 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 is not a marketing summary. It's generated by the agent itself, in real time, showing its work.
You can disagree with it. You can redirect it. You can reject it. And that conversation is preserved — so when you look back, you can see the decisions that were made and why.
The important part is that the record is visible enough for the team to inspect later.
What we don't do
We don't automatically merge anything. We don't auto-run agents on your behalf without a trigger. We don't train on your team's conversations; the current policy is written plainly on the privacy page.
These aren't features we'll add "when you opt in." They're not on the roadmap at all, because we believe they'd undermine the thing that makes hilos worth using: human-authored direction and human-approved output.
What we're building toward
The goal is teams that trust their agents enough to give them more autonomy over time — because they've built that trust incrementally, in a workspace that kept them in the loop the whole way.
Not agents that replace engineers. Agents that make engineers more effective.
The honesty matters for a practical reason too. When the agents do something unexpected — and they will — you need a team that trusts the system enough to investigate rather than panic.
That's the environment we're trying to build. We'd rather do it honestly than pretend we're further along than we are.