what did we give the agent to start the research?
step-by-step instructions
a goal, a photo, a text file
a list of websites to visit
a program we wrote
where did the agent do its work?
all of it on the laptop
files here, model in the cloud
in the browser, like chatgpt
on a university server
8 sub-agents researched at once. who decided that?
our prompt told it to
the agent itself
a setting in the app
it always works that way
the agent asked us which file held the milestones. what does that show?
it could not read the file
it noticed a mismatch
it asks before every step
it had forgotten the prompt
part 2
what is an agent?
who did the work while we watched?
image: ai-generated (gpt-image-2.5-flare)
chatbot vs. agent
a chatbot only answers, while an agent acts on its own.
a working definition
an
AI
agent runs tools in a loop
to achieve a goal.
after simon willison, september 2025
·
anthropic: agents are "
LLM
s using tools based on environmental feedback in a loop"
workflow or agent?
anthropic, building effective agents (2024)
the goal and the model
what we saw:
a few sentences, a photo, a text file.
tools
what we saw:
it listed the folder, downloaded papers, wrote files.
the loop
what we saw:
it read a source, then decided what to do next.
context window
what we saw:
"context 14 %" in the corner of the app.
plan and instructions
what we saw:
"show me your plan. wait for my go."
sub-agents
what we saw:
"research by 8 agents", and nobody asked for that.
all of it together
the model decides, the tools carry it out, and the loop keeps running until the goal is reached.
the harness
the same model in a different harness is a different agent.
models and apps
where does it run?
see also:
open models, an
LLM
on your own laptop.
why now?
the length of tasks agents can finish doubles every few months.
METR, time horizons
·
kwa et al. 2025
· software tasks
an agent acts with your permissions
it can read every file you can read.
it can run every command you can run.
it can publish to the internet.
image: ai-generated (gpt-image-2.5-flare)
we open the loop.
layer by layer, down to the model and back.
image: ai-generated (gpt-image-2.5-flare)