Anatomy of One Agent Run

The trajectory who talked to whom

THE AGENT

USER

HARNESS↻runtime + loop

MODEL (LLM)a stateless, probabilistic function

THE FABRICreached only through tools

composes context: system prompt + tool menu

help ticket

text in: msgs 1-2

text out: get_topology()

get_topology

result → context

text in: msgs 1-4

text out: get_flow_records()

get_flow_records

result: 14 flows dead

text in: msgs 1-6

text out: get_fabric_counters()

get_fabric_counters

result: fabric clean

text in: msgs 1-8

text out: get_host_counters()

get_host_counters

result: host 16: 153 drops

text in: msgs 1-10, 486,286 tokens

text out: the diagnosis

triage response: sick link, L1

The context window what the model sees, every single turn

msg 1system prompt · assembled by the harness

You are a network-investigation assistant for an RDMA
leaf-spine fabric. Help the user diagnose their issue.

msg 2user · help ticket

"We have flows on the AI cluster that just never finish.
They don't error out cleanly, they don't time out fast, they
just dribble — bytes keep moving for a while and then the
flow sits there. Other flows on the same hosts are fine.
Started a couple of weeks ago. No firmware update, no
topology change we're aware of."

msg 3model emits · turn 1

→ get_topology()

msg 4tool result · pasted in by the harness

256-host leaf-spine, bundled burst flow pattern
confidence: high · staleness: fresh

msg 5model emits · turn 2

→ get_flow_records()

msg 6tool result

flows: 266 total · 252 completed · completed p99: 433 µs
14 × "dropped_without_completion"
  sip 0b000001 → dip 0b001001
  sip 0b000101 → dip 0b001001
  sip 0b000201 → dip 0b001001
  ... 11 more, every one → dip 0b001001

msg 7model emits · turn 3

→ get_fabric_counters()

the trap msg 8tool result

every queue on every switch port:
  dropped_packets: 0
one ECN hotspot (if 17: 848 marks)
one transient PFC pause/resume pair (if 34)
~41,900 lines of JSON, abridged

msg 9model emits · turn 4

→ get_host_counters()

the smoking gun msg 10tool result

host   0  11.0.0.1    drop_packets: 23
host   6  11.0.6.1    drop_packets: 20
host  16  11.0.16.1   drop_packets: 153
host 152  11.0.152.1  drop_packets: 1
every other host: 7-23 (baseline noise)

msg 11model emits · turn 5 · final answer

"I've got enough signal to call this."

"Host 16 is ~10× the baseline of every other host.
That's not statistical noise; that's a sick link."

"Reseat / swap the optic / clean the connector on
host 16 first — cheapest fix."

"The fabric itself is healthy. Don't go chasing PFC
tuning or DCQCN parameters... This is L1."

claude-opus-4-7 · input: 486,286 tokens · output: 1,747 tokens

MODEL (LLM) · a stateless, probabilistic function idle · it has no memory of its own, no connection to anything

What you just watched

5model invocations

4tool calls

1growing transcript

final turn: 486,286 tokens in → 1,747 tokens out

The model never touched the fabric. The harness ran four errands and kept one transcript.

Model + prompt + tools + loop. That is the entire machine.

Same scenario with two fewer tools: a confident, wrong ECN story. The menu decided the outcome.