How Agents Work
What an agent is, and the order you build one in.
An AI agent is a software-based digital worker that understands goals, makes decisions, and takes actions to complete tasks. Unlike a chatbot, which mainly generates responses, an agent reasons through problems, reads knowledge you give it, uses tools and integrations, and interacts with external systems — databases, APIs, enterprise applications — with minimal human intervention, while still allowing human oversight where you want it.
This page walks through building one, in the order you actually do it. Each step links to the concept behind it and the screen that configures it.
1. Describe what you want the agent to do
Start with a plain-language description of the task. Be specific about what the agent should accomplish — the more context you give, the less manual configuration you'll do afterwards. You don't need to know which settings to use yet.
What Qore does: The Agent Builder analyses your prompt to pre-fill the agent's name, output format, and an initial skill set. Treat it as a first draft you'll review, not a finished agent.
2. Confirm how the agent identifies and responds
Review the suggested name and description, and choose the Agent Output Format that matches who — or what — receives the responses: Markdown for humans reading in a UI, Json when your application parses the output.
What Qore does: Name and description are metadata for builders; they don't change how the agent behaves. Output format is a contract with whatever consumes the response.
3. Choose the runtime version
New agents are pinned to the latest executor version your tenant supports. Existing agents stay on the version they were created with unless you change it deliberately — the platform won't force an upgrade. Move to a newer version when you want a capability it introduces.
What Qore does: The executor changes over time — new versions add capabilities, fix issues, and sometimes deprecate older behaviors. Switching versions can change agent behavior and the set of features available, so re-test after you do.
4. Set how the agent thinks
Three levers, all on one screen, all with sensible defaults:
Agent Capability (default
Auto) — how deeply the agent thinks about each request.Autolets the platform decide per request, which is right for most agents.Response Mode (default
Advanced) — the speed-versus-quality trade-off.Execution Model and Planning Model — which model runs the agent, and which one plans its work. They can differ.
What Qore does: Planning happens once per request; execution may run many tool calls. That's why the two models are separate — a common pattern is a heavier model for planning and a lighter one for execution.
→ Steering & Reasoning · Agent Behavior
5. Decide what the agent remembers
Memory (default off) lets the agent remember what an individual user has asked about, preferred, or corrected, across conversations. Turn it on for assistant-style agents in a chat UI where users log in. Leave it off for headless agents called by API with no persistent user identity — there it adds noise without benefit.
FAQ memory (default on) lets the agent reuse a stored answer when a question matches a known FAQ.
Confirm the scope before enabling memory on an agent handling sensitive material. Who can see what a user told which agent is worth answering explicitly in a regulated environment — check with your platform administrator.
→ Memory & Learning · Agent Behavior
6. Remove capabilities the agent doesn't need
The platform attaches a set of skills to every agent by default. Read that list and disable anything your agent won't use — keeping unused skills active widens the range of things the agent can do wrong.
What Qore does: Platform-supplied skills are attached automatically and can be individually disabled without affecting the ones that remain.
→ Skills & Tools · Skills & Capabilities
7. Connect the tools and data the agent needs
For each external system the agent must reach, attach the right kind of skill: an API call for REST endpoints, an MCP server for tool platforms, a RAG container for document search. Configure endpoints, authentication, and schemas for each.
→ Adding a Skill to the Agent · MCP Tools · Knowledge & Context
8. Decide how much scrutiny the agent applies to itself
Enable COT Validation (LLM Judge) when a wrong answer has real consequences — approving a document, triggering a payment, anything hard to reverse. Skip it for conversational or high-volume cases where speed matters.
What Qore does: COT Validation uses a second, different model to review the output before the user sees it — so every response waits for two models instead of one. Disabled by default.
Guardrails are the broader mechanism: validators that inspect content at four points — the inbound request, each tool call before it runs, each tool result, and the final response.
→ Guardrails & Validation · Guardrails (Beta)
9. Test whether it behaves as intended
Use Try Agent to run sample inputs that represent real usage. Open the execution trace to see how the agent interpreted your instructions, which skills it called, and how long each step took. If something's off — wrong skill, unexpected output — adjust and re-test before moving on.
What Qore does: The execution trace is for builders. End users see only the final response.
10. Go live, and keep it improving
Once the agent performs on your test cases, deploy it. Then review what it actually does in production: approve corrections that represent genuine improvements, reject anything incorrect or exploitable, and iterate as edge cases emerge.
What Qore does: Learning cases require manual approval before they enter agent memory. Auto-approval is intentionally disabled, so a malicious or mistaken input cannot quietly corrupt the agent's behavior.
→ Deploy · Monitor · Human-in-the-Loop
The loop closes: what you learn in step 10 becomes a change in steps 4 through 7. For the ideas behind each step, see Agent Foundations.
Last updated