Everyone's still trying to write better prompts.
you
write every step
every step
prompt
expense.ts, 3 cases
claude
one shot, no memory
output
a diff you have to read
YOU ARE THE LOOP
everyone's still trying to write better prompts
The people who shipped Claude code stopped prompting months ago.
the people who shipped Claude code
stopped prompting months ago
They upgraded themselves.
First to loop engineering.
GOAL YOU WROTE
every file has a test
npm test passes
stop after 20 turns
claude
acts until the goal is true
verifier
a separate model
goal met?
done
THE LOOP RUNS ITSELF
they upgraded themselves first to loop
engineering
Now to graph engineering.
ORCHESTRATION
scan src/
finds the work
one per file
agent
cart.ts
branch ad
agent
user.ts
branch ad
agent
auth.ts
branch ad
agent
api.ts
branch ad
merge
one result
e-run until nothing new
now to graph engineering
let me show you inside Claude Code
I built a tiny expense tracker
it works but it has zero tests
level 1 prompt engineering
write tests for ExpenseForm.tsx
Cover:
- adding an expense
- deleting an expense
- rejecting negative amounts
I write the instructions myself
write tests for expense form dot t
sx
level 1 prompt engineering
it works until you have 10 more files
now you're just copy pasting prompts
level two loop engineering
this is where things changed
as Boris Cherny the creator of Claude Code described
he stopped thinking about prompting
He stopped thinking about prompting.
he stopped thinking about prompting
and started building loops around Claude
He started building loops around
Claude.
instead of telling Claude every step
STOPPED
STARTED
LOOP ENGINEERING
you define the finish line
/goal
Every file under src has a matching test.
npm test passes.
stop after 20 turns.
GOAL, VERIFICATION, STOPPING CONDITION
every file under s R C has a matching test
a N p m test passes stop after 20 turns
Instead of telling Claude every step,
you define the finish line.
THE FINISH LINE
every file has a test
npm test passes
stop after 20 turns
claude
acts until the goal is true
verifier
a separate model
goal met?
done
GOAL, VERIFICATION, STOPPING CONDITION
now you're defining the goal
the verification and the stopping condition
Claude keeps working until the goal is true
level 3 graph engineering
run test cases
run test cases
this is where things get interesting
as Peter Steinberger asked:
Post
Peter Steinberger
@steipesa
Are we still talking loops or did we shift to graphs yet?
8:34 AM Jul 18, 2026 2.6M Views
1.1K
979
7.4K
2.6K
this is where things get interesting
as Peter Steinberger asked
LOOPS
GRAPHS
are we still talking loops
Because now you're not controlling one agent's loop.
or did we shift to graphs yet
GRAPH ENGINEERING
You're not controlling one agent's loop.
writer
cart.ts
planner
splits the work
writer
user.ts
reviewer
attacks it
validator
die at catch
merge
one result
because now you're not controlling one agent's loop
You're designing a system of agents.
src/
12 files to cover
WORKFLOW SCRIPT
writer
cart.ts
writer
user.ts
writer
auth.ts
writer
api.ts
reviewer
mutates
reviewer
mutates
reviewer
mutates
reviewer
mutates
validator
must fail
validator
must fail
validator
must fail
validator
must fail
reviewer validators running in parallel
ALL FOUR COLUMNS RUN AT THE SAME TIME
reviewer validators running in parallel
That's what makes graph engineering different.
merge
1 verified
report weak
1 test caught nothing
Prompt engineering scales you.
prompt engineering scales you
Loop engineering scales one agent.
loop engineering scales one agent
Graph engineering scales an organization of agents.
graph engineering scales
an organization of loop of agents
and if you're still wondering what IDE I'm using
this is DevSwarm
it lets you run multiple projects in parallel
each with multiple isolated Git branches
and its own IDE so your agents never step on
each other
try it for free at devswarm.ai
try it for free at devswarm.ai