AI agents with memory and tools.
A chat session forgets everything the moment you close it. An AGNT agent is a persistent worker — it has its own memory, its own tool permissions, its own model, and a record of everything it has done before.
Configuration, not conversation
Every agent is a stored, versioned object. You define what it can touch and what it costs, and that definition persists across sessions, restarts, and machines.
Scope and purpose
Name, description, icon, and category. An agent is built for a job — a research agent, a release agent, a support triage agent — not as a general chat window.
Assigned tools
You grant an explicit tool set and workflow set. An agent cannot reach for something it was never given, which makes its blast radius a design decision rather than a surprise.
Credit limits
Each agent carries a credit ceiling and tracks consumption against it. Autonomous work cannot quietly run up an unbounded bill.
Agents that compound instead of resetting
Memory is typed, not a single undifferentiated blob. The type determines how a memory is retrieved and how heavily it is weighted — so a correction you made once does not get lost among a hundred trivia facts.
| Memory type | What it captures |
|---|---|
| fact | Stable truths about you, your stack, or your business |
| preference | How you like things done — style, format, tone, defaults |
| correction | Something the agent got wrong and you fixed, so it does not recur |
| context | Background on a project, codebase, or ongoing situation |
| pattern | An approach that worked and is worth repeating |
| tool_insight | Hard-won knowledge about a specific tool's quirks and failure modes |
| workflow_insight | Optimizations discovered while running a workflow |
| prompt_guidance | Prompt phrasing that reliably produces better results |
Memory is stored locally alongside your execution history. It is yours, it is inspectable, and it never gets shipped to a vendor for training.
Every run leaves evidence
Full execution traces
An agent run is not a black box. Each execution records what was asked, what was answered, and every tool call in between.
- The exact prompt that started the run
- Every tool call with its input, output, and any error
- Timestamps, token usage, and cost per step
- The final response, stored verbatim
Searchable history
Traces are indexed, so "what did we do about that bug last week" is a query rather than an archaeology project.
- Full-text search across every past run
- Success rate tracked per agent over time
- Reconstruct any run step by step, months later
- Receipts stored locally in SQLite, not in a vendor's cloud