HarnessX

The deciding trait is that turning a coding agent into a research agent is a configuration change rather than a rewrite. Behaviours snap together, so the second agent you build costs a fraction of the first, and the third is nearly free. Anyone who has copied a loop between two projects and watched them drift will recognise the appeal.

You are the wrong buyer if you need one working agent today, because assembling is still assembling. Pick it when you are building a family of them. Pick a finished tool when you are building one.

The risky part is the part being advertised. A meta layer watches the agent run and proposes different arrangements of its own behaviour, which means the thing you tested on Monday is not necessarily the thing running on Thursday. No acceptance criterion is documented, no rollback is described, and no ceiling is placed on how far a proposal may drift from what you configured.

What it does right is name the tiers of isolation. Local, container and hosted are three separate documented backends rather than one setting that means different things.

1. The central expression puts provider routing and per-role model assignment on one side and the behaviour pipeline on the other, which is a real separation rather than a naming convention.

2. Behaviours compose with an operator, so an agent's definition is an expression that can be read left to right, and two agents can be compared by comparing their expressions.

3. Runs produce reward-annotated records intended to feed fine-tuning, which is a coherent ambition. No result from that pipeline is published, so the loop is described rather than demonstrated.

454 stars, a small research-flavoured organisation, and no price anywhere. What is being built here is closer to a lab's tooling than to a company's product, and that is not a criticism, it is a description of where the effort goes: into the idea rather than into the funnel.

Moat: the research, briefly, until it is published or copied. Likely acquirer: a lab hiring the team, which is how this shape of project usually resolves. Position: watch it, borrow the ideas, and do not put it under anything that has to run next quarter.

One flag runs a single prompt and exits, so this can sit in a pipeline and be scheduled like anything else we operate. That is more than most of this category offers and it is the reason I read the rest of the row.

The rest is a procurement exercise. One of the isolation backends is a third-party hosted service, which means our code reaches a vendor I have not reviewed unless we pin the local option. No SSO, no audit log, no retention policy. Approved with conditions: the hosted backend disabled by default.

MIT, and it installs into a virtual environment I control rather than into a global path. Every behaviour is an ordinary Python object, so extending it means writing a class rather than filing a feature request, and reading it means reading Python instead of a plugin manifest.

The gap is the tool layer. Capability comes from an internal registry, and there is no MCP client, so servers I already run have to be re-exposed as native tools before they exist here. That is a day of work I did not want to spend.