I keep seeing the term and every explanation I find is either three words long or a security paper. I am trying to explain to non technical colleagues why we cannot just let an assistant read the shared inbox and act on…
Something changed this autumn and it is bigger than a smarter model. The new generation of assistants can fill in forms, update CRM records, work in spreadsheets and test software on their own. Microsoft rebuilt Copilot…
Everybody recommends MCP servers and nobody says how big they are, so I counted. I pulled a year of posts from the agent and MCP communities, extracted every GitHub link in the bodies, and checked all 552 of them…
An agent went into a loop with real side effects and I discovered there was nothing to press. What I have in place now, in order of how reliably it works. Things that stop it A hard step and token ceiling per run. Not a…
An agent with a handful of tools. One of them hit an authentication error and returned an error object describing the failure. The agent did not stop, did not retry, and did not mention it. It carried on and produced a…
A recurring theme in recent launches is approval agents that ask before acting, spending allowances rather than open credentials, setups aimed at one person supervising several workers. The underlying problem is real…
I have a workflow triggered by incoming messages from a messaging platform. On the vendor's hosted version it works immediately. I exported it, imported it into my self hosted instance with the same credentials, and the…
Three separate products landed this month that are, in one way or another, a browser built for AI agents rather than for people one positioned as the browser for coding assistants, one running on edge workers, one aimed…
In the past few weeks I have seen an MCP server announced for a brand asset service, one for keeping prompt context portable between tools, one for local semantic search. It has become a standard thing to ship alongside…
First agent I have built, so possibly a naive question. I give it a task with nine clear steps in the prompt fetch, validate, transform, write four files, update an index, report. It reliably does the first three, then…
Support agent over our own documentation. One user question something about the refund window and the loop made 38 calls to before hitting and giving up with an apology. Trace is attached. Nine of the calls are byte…
Internal agent for our ops team. It has grown to 40 tools across four domains deploys, tickets, the data warehouse, and a pile of one off scripts. Counting the definitions alone, that is about 11,400 tokens attached to…
Tool takes . Maybe one call in six arrives as instead of , and my Zod parse throws, and the agent then spends two more steps apologising and trying again. The schema comes out of . Looking at the emitted JSON I see the…
Same server, same registration code. Over stdio the client lists all 12 tools and they work. Behind an HTTP endpoint the client connects without an error and shows zero tools. What I can see server side: the request…
The agent writes and runs shell commands against a checkout of our repo tests, migrations, the odd script. About 200 runs a day, all triggered by me, no public input anywhere in the path yet. Right now it runs on my…