LenzIndependent, multi-model fact-checking API for AI workflows
Promoted
Cartha — Agent Operations Platform
Ops layer for AI agent fleets: traces, memory, hard budgets
5 followers
Ops layer for AI agent fleets: traces, memory, hard budgets
5 followers
Most teams get AI agent traces. Almost nobody gets governance. Cartha is the ops layer for agent fleets: full run timelines, scoped memory enforced on the server (user/agent/team/org), hard spend budgets that fail closed, attenuated multi-agent delegation, policies, and human-in-the-loop. Instrument in a few lines (Python SDK): @trace, @tool, wrap_openai, then see agents, costs, and memory in one dashboard. Not another chatbot builder. You bring the agents; we keep them accountable.
One thing I'd love to see is a way to replay a specific agent run step-by-step with the ability to fork it from any intermediate decision point. That would make debugging weird multi-agent handoffs so much easier, especially when something goes sideways three layers deep and you want to test a different policy branch without rerunning the whole thing.
@trkan7zmg Thanks, this is a great callout, and exactly the pain we keep hearing from multi-agent setups.
Today Cartha already helps on the forensics side of that problem:
• full step timelines for nested agents
• policy replay on a past run (“would this still be allowed under current rules?”)
• run-diff between two runs
• time-travel memory (as_of) so you can see what the agent knew at a past decision
What you’re describing is the next level: fork from an intermediate step, change a policy / tool / memory branch, and continue without replaying the whole graph from t=0. That’s on our radar as “branch / what-if from step N”, especially for handoffs three layers deep.
If you’re open to it, I’d love a concrete example of a handoff that went sideways for you (even rough). Happy to share how far the current replay path gets you today and what a first fork UX would look like.
Live product: cartha.in (https://cartha.in) — feedback always welcome.
The hard spend budgets that fail closed are a great touch. One thing that would make this even more useful for us is a way to set per-tenant or per-project policy profiles, so a staging environment can run loose while production stays locked down without duplicating config.
@faruk1723009 Appreciate that, fail-closed budgets are exactly for the “oops, the agent looped overnight” case.
On per-tenant / per-project policy profiles (staging loose, prod locked, no config copy-paste): totally agree, and it’s a common ask once teams leave the single-env demo.
Today you can approximate it without forking the product:
• separate orgs (or API keys) for staging vs prod, each with its own policies/budgets
• or separate teams under one org, with different agents + policy sets
What you don’t get yet is a first class named environment profile (“staging / production”) you can attach to a project and switch with one label, same policies as code, different strictness per env, no duplicate dashboard clicking.
That’s high on the list for multi-env teams: policy profiles / environments bound to project or tenant, so prod stays tight and staging stays exploratory without cloning config by hand.
If you want to jam on the shape (profiles vs tags vs separate orgs), happy to, and if you’re testing Cartha already, tell me whether you think in “projects,” “tenants,” or “deployments.” That changes the UX a lot.