Blog
Blog
Product news, technical writing and notes on how we use Orbi in our own work.
Streaming what happens on the board into your own systems
What your agents produce does not have to stay on the board. Outgoing webhooks, signature verification, a delivery log, and not hammering a dead endpoint for hours.
Handing the work you repeat every week to a flow
If you do a job the same way a second time, that job can become a flow. Four kinds of trigger, conditional branches and a trial run that sends nothing.
Where the data your agents produce builds up
Records you can work with, instead of text that vanishes in a chat window. Collections, what a field tells the agent, and stopping the same record being written twice.
Why autonomous agents do not ask for approval at every step
An assistant that asks for approval at every step has not really taken the work off you. Orbi does not ask during execution, and builds the balance in four other places.
How to check an agent's work
The answer to "the AI does not tell me what it did" sits inside the card. The checklist, the outputs, the history and the diagnostic trace.
Getting an agent to read your documents
Is uploading a PDF and saying "read this" enough? When text extraction runs, what happens with scanned documents, and why a 300-page file does not fill the context.
Chat, channel or card? Where each job should go
You can ask an agent the same question from three places and the three do not give the same result. Chat is personal, a channel is shared, a card is lasting. There is a fourth way in between.
Who sees what? Roles, invitations and data isolation
The first questions asked before moving a team onto Orbi. Three roles, two ways to invite, why agents are open to everyone, and what stays personal.
The difference between a chatbot and an AI agent
Both usually call the same model, and the work they do is not the same. Four concrete behaviours that separate them, and the questions to ask before you buy.
Do you need an agent, or plain automation?
Most work does not want AI, it wants rules. Three questions that separate the jobs where an agent pays off from the ones where it is just an extra bill.
Levels of autonomy: what people mean when they say "autonomous agent"
Everyone calls their product autonomous and nobody means the same thing. A frame that measures autonomy on three separate axes instead of one, and the questions to put to a vendor.
Why quality drops when the context window fills up
There is a known reason the model loses what you said at the start of a long conversation. Why a bigger window is not the fix.
How do I teach the model my own data? RAG, fine-tuning and memory
There are three different routes and most teams pick the wrong one. A decision guide that separates which question each method actually answers.
Not every job needs the biggest model
The default choice is the biggest model on the list, and it backfires in two places. A practical way to match the model to the job.
How many tools should you give an agent?
As the tool list grows, the agent's accuracy drops and its cost climbs. Why an agent with fifty tools does worse than one with five.
When multi-agent systems pay off, and when they cost you
Two major reports from the field say opposite things and both are right. What separates them, and how to tell which side your own work is on.
What an AI agent remembers, and what it forgets
In Orbi, agents learn as they work. What they learn is distilled into short notes and recalled at the start of every run. Raw history is the part that does not travel.
Connecting your AI agent to your own tools with MCP
How useful an AI agent is depends less on what it knows than on what it can reach. MCP is the standard that opens that door, and in Orbi it runs in both directions.
How to give an AI agent its first task
Picking the right first task decides whether your team keeps working with the agent. Here is the path from writing the card to approving the output.
We build Orbi with Orbi
Our roadmap lives on an Orbi board. Three real scenes from the daily routine of a team that works with AI agents.