Obliq Agents
An Obliq agent binds knowledge, instructions, a model, memory, tools, and policy. Every run can be inspected and replayed.
Not a chatbot builder. A system for agents that know, act, and improve.
Who agents are for
Use agents when the job has boundaries: a support workflow that may open a ticket, a research dispatch that may only read approved sources, an internal assistant that must show which document it used. If those boundaries live only in a prompt, they disappear the moment the model or the tool changes.
How an agent runs
A request hits the agent. Obliq assembles context from the knowledge it is allowed to see, calls the assigned model, and invokes a tool only if policy permits. The result, the sources, and the tool payloads are stored on the execution. You can replay that input with another model or tighter instructions and compare the two trails.
What an agent contains
The sources this agent may retrieve, scoped so it cannot wander into unrelated corpora. Each retrieval is attributable in the execution trail.
Persona, constraints, and the output shape the agent must follow. Instructions stay with the agent when you change the model behind it.
A provider-agnostic endpoint assignment. Bring your own key or a private endpoint; the rest of the agent definition does not embed a vendor SDK.
Session and episodic working state the agent can read on later turns. Memory is part of the run record, not a hidden side channel.
MCP servers, APIs, webhooks, and functions the agent is allowed to call. Grants are explicit; unlisted tools are not available.
Permissions and human-in-the-loop gates for sensitive actions. A policy can require confirmation before a tool fires.
How it fits the system
Agents sit between knowledge and action. Knowledge supplies what the agent may say is true. Tools supply what it may do. Observability supplies the proof. Evaluation scores groundedness, instruction adherence, and tool reliability so the next change is based on a run, not a vibe.
Questions
What is an Obliq agent?
An Obliq agent is a production runtime unit: scoped knowledge, instructions, a model endpoint, memory, tools, and policies. Each execution leaves a trail you can inspect and replay.
How is this different from a chatbot builder?
A chatbot builder wraps a conversation UI around one model. An Obliq agent is defined by what it may know, which systems it may call, and how the run is recorded. The chat surface is optional; the runtime is not.
Can I change the model without rewriting the agent?
Yes. The model is an assigned endpoint. Instructions, tool grants, and knowledge scope stay in the agent definition, so you can route the same task to another provider or a private endpoint and compare the traces.