Introducing VexaAgent — production agents you can actually trust

Agent frameworks have a credibility problem. Every framework demo works. Almost no agent built on those frameworks survives contact with production.

The failure modes are predictable:

• ›The agent hallucinates a tool call.
• ›The agent forgets it already retrieved the answer two steps ago.
• ›The agent burns tokens looping until a budget cap stops it.
• ›The agent confidently produces an output with no traceable evidence.

We've been building infrastructure that makes those failure modes harder to hit — ContextAPI for token-aware prompt construction, DataAPI for grounded retrieval — and we kept hearing the same question from customers using both: Can you give us the loop too?

Today we're opening the private beta of VexaAgent.

What it is

VexaAgent is a production agent runtime. You describe a goal, give it tools, point it at a DataAPI tenant. It plans, retrieves, composes, and verifies in a tight loop — every step grounded, every output cited, every token accounted for.

The runtime is built on three Cosavu primitives:

• ›ContextAPI wraps every Vexa-1 call. Prompts are typed, compressed, and budget-enforced before they reach the model.
• ›DataAPI is the default retrieval tool. Grounding isn't optional — the agent can't compose a final response without at least one DataAPI call.
• ›Vexa-1 is the underlying reasoning model. Trained specifically for tool-use and multi-step planning.

What's different

We made three opinionated decisions that distinguish VexaAgent from the agent frameworks we'd seen.

Grounding is required, not optional. Every final response cites at least one retrieved document. The runtime blocks outputs that don't cite. This eliminates the "confident hallucination" failure mode at the architectural level.

The loop has a budget, not a step count. Instead of "max 10 steps", the runtime tracks token spend and tool-call cost. The agent decides when to stop based on remaining budget. This eliminates the "ran out of steps mid-task" failure mode.

State is keyed and persistent. Every intermediate reasoning step, every tool output, every retrieval result is logged and queryable. You can replay any agent run end-to-end. This eliminates the "what did the agent actually do" debugging problem.

What it's good at

We've been running closed beta with eight customers since January. The workloads where VexaAgent has been a clear win:

• ›Customer support copilots. Drafting responses grounded in policy docs and ticket history.
• ›Internal research. Multi-source synthesis across knowledge bases with full citations.
• ›Triage and routing. Reading incoming requests, classifying, dispatching to the right team.

The workloads where it's less mature: creative generation, open-ended ideation, anything where grounding isn't a meaningful constraint.

Getting access

VexaAgent is invite-only for now. We're onboarding teams that are already on ContextAPI or DataAPI and that have a specific production workflow in mind. If that's you, request access from your account team or via the contact page.

We'll open the public beta later this quarter once we're confident the runtime handles the edge cases that show up at 10× our current load.