First Look At The Avi Agent System

The only limit is your imagination

Your personal AI workforce. Not a chatbot. A team of autonomous agents that research, monitor, analyze, and deliver, on a schedule, around the clock, while you sleep.

What Is Avi?

Avi is a self-hosted autonomous agent platform. You define what you want done. Your agents just go and do it, gathering data from dozens of live sources, analysing findings, writing reports, sending you alerts, and getting smarter every cycle.

Unlike typical AI tools that require you to prompt them every time and forget everything the moment you close the window, Avi agents have persistent memory, self-written playbooks, reputations that grow over time, and a fully visual control center where you manage everything with a click.

Agents don’t just work on your tasks, they evolve. They find better ways to work on your task, build their own tools, and think on their own.

Here are just a few of the panels

Watch Them Work

A live event feed shows you exactly what is happening across your entire agent network, which agents ran, what they found, whether any alerts fired, and when approvals are needed. It is like watching a team of analysts working in real time.

When you have multiple agents working together in a chain, for example, one agent researching raw data and another compiling it into a report, the Pipeline Canvas visualizes their interaction. You can see the entire workflow at once: which agent feeds into which, and where in the chain of execution currently is.

Watch The Intelligence Feed

A curated live feed of intelligence items gathered by all your agents, surfaced in one unified place. New findings from any source, news, security alerts, market data, academic papers, appear here as they are discovered.

There is also a live display showing which tools your agents are using, how often, and whether any are hitting rate limits or failures. Useful for understanding what data sources your agents are actively pulling from.

Everything At Your Fingertips

Press a single shortcut to open the Command Palette, a fast-access launcher for any action in the system: triggering an agent, searching the knowledge graph, opening settings, viewing a log. Keyboard-first power users will feel right at home.

You can also run a full-text search across everything, journal entries, agent outputs, learnings, knowledge graph entities. Find any piece of intelligence your agents have ever gathered, instantly.

The Individual Agent Workspace

Click any agent in the sidebar to open its dedicated workspace. Each agent has its own tab-based control panel with every detail you could need.

Overview Tab

Your at-a-glance command view for an agent:

• Current Focus, what the agent is actively working on right now

• Goals with progress bars, color-coded green (on track), amber (in progress), red (struggling)

• Reputation Score, a 0–100 score computed from the agent’s grade average and consistency over time

• Dimensional Self-Grades, the agent rates its own output across five axes: Depth, Relevance, Novelty, Actionability, and Quality. Visualized as bar charts so you can see at a glance where it excels or needs help

• Growth Narrative, a plain-English summary like *”Cycle 47 agent | avg grade 7.8/10 | trending ↑ | 12 👍 2 👎”

• Expertise Tag, badges showing the domains the agent has developed competency in

• Cycle Type, whether the last run was a deep research cycle, a quick maintenance scan, or a deliberate skip

• Confidence, how confident the agent was in its own findings (self-reported)

• Blocked / Needs Input, flagged prominently if the agent is waiting for you

The Agent Persona

Each agent has its own persona, its name, mission, operating principles, voice, and approach. Think of this as the agent’s job description and personality definition. Fully editable at any time.

And each agent writes a journal

After every cycle, the agent writes a structured entry covering what it did, what it found, what surprised it, and what it plans to focus on next. You can toggle between recent entries and the complete archive going back to day one. If the agent didn’t write a journal entry on its own, the system auto-generates one so you are never left in the dark.

The agent that manages its own goals

Agents have goals. Goals you create, and goals that it creates itself. Each goal shows a progress estimate. Agents also evolve and break goals down into sub-goals that help it be more efficient in its planning and task management.

Agents have memory

Avi agents have an advance memory system not seen in any other agent system. They have a custom built vector store which links facts, memories, locations, events, people, stories, articles, and everything it knows, together in a complete graph view that you can watch evolve live.

Your agents get smarter over time.

It also maintains a curated, distilled and personal knowledge base the agent builds over time. When an agent discovers something important that should inform all future work, a key insight, a pattern, a lesson learned, it records it here. This is not raw notes; it is refined intelligence that persists permanently.

And it has tactics. This is the agent’s self-written playbook. Over many cycles, agents notice what works and what doesn’t. They record rules like ”leads with a data table, gets higher grades” or ”avoid generic summaries, creator consistently thumbs-down those.” Agents read this playbook at the start of every cycle and adapt their approach accordingly. You never have to teach them explicitly, they figure it out themselves.

It also maintains a scratchpad. This is an agent’s working memory for a given cycle. Rough notes, intermediate findings, half-formed hypotheses. This resets each cycle and gives you a transparent window into the agent’s current thinking process.

The Secret Source, Advanced Planning

When an agent operates in Planning Mode, it generates a structured execution plan at the start of each cycle before taking any action. This is they key to Avi’s evolution. When agents are not in planning mode, they follow your instructions, they get routine tasks done, they are effectively an intelligent n8n or Zapier.

In planning mode, they know the goal, they work on the best way to do it, and they evolve to do it better over time. This is the side of agents you will not have seen before. It truly amazes me every day.

Look out for the release tomorrow.

The only limit is your imagination