Raven

The deciding trait is memory that survives the session. Most agents forget you between windows, so you re-explain the project every morning like a new contractor arrived. This one carries long-term memory across sessions, and after a fortnight the difference is not a feature comparison, it is whether you have to repeat yourself.

The catch is that memory cuts both ways: a wrong belief also persists, and you will spend time correcting one. Pick it if continuity is what you want. Pick something stateless if you prefer a predictable blank page.

Two moving surfaces multiply. The vendor's own warning is that configuration will shift under you, and the skill component is designed to evolve the instructions it retrieves, so today's behaviour is produced by material neither you nor the maintainers wrote in full. When something regresses, the question of what changed has two candidate answers and no obvious way to separate them.

What it does right is keeping the reasoning path recorded rather than discarded, which at least makes the postmortem possible.

1. Explicit budgeting is the correct treatment of a finite window. Deciding in advance what share belongs to history, tools and instructions produces failures you can reason about, whereas implicit truncation produces a silent loss whose cause is invisible in the transcript. 2. Shipping an evaluation loop inside the harness is unusual and welcome, since it makes regression a measurable event rather than an impression.

2. No results from that loop are published, so the apparatus is documented and the outcomes remain unreported.

Attention this size is an asset and it is not revenue. The architecture makes monetisation awkward on purpose, because everything that would normally be the hosted product runs on the user's own machine, and the usual answer to that is a control plane the audience explicitly does not want.

Moat: the memory layer, if the format becomes something users refuse to abandon. Likely path: a hosted team edition, or an acqui-hire by a platform that wants the agent runtime. Position: keep an eye on it, and do not build a company process on a pre-revenue local runtime.

The messaging surface is the whole conversation with my security team. Twelve integrations means twelve destinations where a repository excerpt can be pasted into a consumer platform, with no central policy, no directory and no retention setting to point at during an audit. That is a questionnaire I cannot complete.

It does run headless, which is the one thing that would make it a pipeline component worth measuring. The licence costs nothing across sixty seats and the governance cost is the real number. Not yet.

Local tracing by default is the detail that earns my trust. Every other product treats observability as a reason to require an account; here the reasoning path lands on my filesystem and stays there, which means I can inspect a bad run without agreeing to terms of service to do it.

Weights can be mine through two local serving options named in the documentation, the licence keeps a fork viable, and the only thing I miss is MCP, which is absent at both ends and would have saved me writing adapters.