An agent is a loop: think → pick a tool → observe → repeat until the goal is met or a budget runs out.

Three execution modes

The agent runtime picks a path per provider:

Pure chat — no tools used, just streaming answers
Native tool calling — used when the provider supports it (Claude, OpenAI / OpenAI-compatible, Gemini)
ReAct fallback — text-based tool calls, used for Ollama (which doesn't expose native tools)

Built-in tools

Six tools today:

Tool	Action	Bounds
`list_tables`	List tables in the database	—
`get_columns`	Get a table's columns / indexes / FKs	—
`execute_query`	Run a read-only SQL query	≤ 100 rows; masking applied
`get_sample_data`	Sample rows from a table	≤ 100 rows (default 20); masking applied
`execute_statement`	Run a SQL statement	Permission-gated
`get_collection_info`	Vector collection metadata only	No similarity search

Masking is enforced server-side before results return to the model.

Permission tiers

Pick one when the agent starts:

Tier	Allows
Read-only	Tool reads + read-only queries
Read-write	Adds `execute_statement` (blocks DELETE / DROP / TRUNCATE)
Schema design	Adds CREATE / ALTER
Full control	Everything

Budgets

Max LLM rounds: 10 (fixed in the agent config)
Tool timeout: 30 s read / 60 s write
Consecutive failures: stops after 3

There is no separate token budget field. Provider max_tokens is currently hardcoded to 4096 in several call sites.

Safety

Panic recovery + duplicate-call dedup built into the run loop
Tool errors back to the model so it can correct itself
Destructive operations are denied based on the permission tier — even with Full Control, the safety checker still reviews each tool call's SQL

Agent runs consume the daily AI quota (license_consume_ai). Free: 5/day; Pro: unlimited. Note: long-term memory auto-extraction is Pro-only — Free plans can read existing memories and manually pin facts (up to 50), but new facts aren't auto-saved from chat.

Not yet built

HTTP / file / shell tool executors — only DatabaseToolExecutor exists today
Agent-level token budget (only max_rounds currently)
Vector similarity-search tool (only get_collection_info exists)

AI Agents