skip to content
behindthestack
lab
about
rss
newsletter ↗
02
agents / llm
Models wired to tools, and where the wiring fails quietly.
Field notes from running agents outside a demo.
channel
model
tools
memory · knowledge graph
plan → act → observe
chat
llm
agent
mcp server
web search
files
loop × n
context ≤ 64,000 tok
1
empty reply,
no error raised
2
mcp token owned
by the wrong user
3
two deployments,
one bot token
fig. 02 — agent loop, one request
● = message in flight
02.1
A self-hosted assistant on a 64K budget
3 min
02.2
MCP: sampling, tools, and who owns the token
3 min
02.3
Recon fleets before building anything
3 min
02.4
A public Q&A bot, deployed twice by accident
3 min