Same bug, four minds: agent behavioral divergence

Edge Delta

5 min read Original article ↗

01 · The shared starting point

Everyone found the same bug

A 1-based severity rank indexing a 0-based, length-3 slice. Introduced by a deploy about 1.5 hours before the alert; latent until a ground_stop corridor hit the handler.

SkyNetOps pattern anomaly

Critical

Fired 2026-07-10 20:19:37 UTC onservice.name:api-gateway, broadcast simultaneously to all four agent channels.

ed.tag:"k8s-eks-skynetops" for
service.name:api-gateway
> 1 count within the last 15m0s

Impact was partial, not total: request volume stayed steady at ~68% of log share. Only /api/route-flows calls touching a ground-stop corridor died mid-flight.

// apps/api-gateway/main.go · handleRouteFlows
sevRank := map[string]int{
  "delay":        1,
  "ground_delay": 2,
  "ground_stop":  3, // 1-based
}
sevHist := make([]int, 3)      // valid idx: 0, 1, 2
sevHist[sevRank[c.Severity]]++ // ground_stop → sevHist[3] 💥
http: panic serving: runtime error: index out of range [3] with length 3
main.(*server).handleRouteFlows

The fix all four converged on: zero-base the rank map. One line.

02 · The race, second by second

Only one closed the loop

The same monitor paged all four teams at 20:19:37 UTC. Every event below is reconstructed from the thread transcripts. All times UTC.

telemetry reconfix workPR openedproductiondead end

▲ alert fired 20:19:37.451 UTC · all four teams paged at once20:2020:2120:2220:2320:2420:2520:2620:2720:28

OpenAI · SRE · 2 sessionsclosed the loop in production

Bad commit traced

20:22:32

Clone + code read

20:23:05

Anthropic · SRE · 1 sessionstopped at "PR opened" by choice

Google · SRE → SWE relay · 2 sessionsstopped at "PR opened" by choice

Clone + bug found

20:22:03

Handoff SRE → SWE

20:23:11

404 on wrong repo

20:23:18

Anthropic · SRE · 1 sessionstopped at "PR opened" by choice

Patterns + events

20:20:20

Second repo sweep

20:23:24

The large marker on each track is the last thing the team did. PR #160 was merged ten seconds after it opened. Three near-identical PRs were still open after it shipped.

03 · Behavioral fingerprints

Four models, four personalities

Identical inputs: same page, same telemetry, same repository, same tools. The run sheets read like four different engineers.

01

GPT 5.5

OpenAI · SRE · 2 sessionsclosed the loop

Time to PR
4:53

Tool calls
64

Tokens
1.30M

PR
#160

“Before summarizing, I need current live confirmation on whether the remediation has landed and whether the alert condition has actually cleared.”
  • -Treated the page as an outage to close, not a ticket to file.
  • -Retrieved its prior false-positive memories about api-gateway and overrode them against live data.
  • -Merged its own PR #160, then polled deployment runs until the panic pattern stopped.

02

Opus 4.8

Anthropic · SRE · 1 sessionstopped at PR

Time to PR
4:29fastest

Tool calls
21leanest

Tokens
1.07Mcheapest

PR
#159

“Recommend fast-tracking review/merge and deploy, then confirming the panic pattern drops to zero.”
  • -Leanest run of the field: 21 tool calls, no wasted motion.
  • -Reconstructed the full goroutine stack to pinpoint the exact line, and quantified the blast radius (68% of traffic unaffected).
  • -Handed merge and deploy to humans with a fast-track recommendation instead of finishing itself.

03

Gemini 3.5 Flash

Google · SRE → SWE relay · 2 sessionsstopped at PR

Time to PR
5:25

Tool calls
44

Tokens
2.21Mpriciest

PR
#161

“Cloning the ozan-skynetops/skynetops repository to work on the correct repository codebase.”
  • -Only team to run go vet and the test suite before opening a PR.
  • -The SRE-to-SWE relay bought clean role separation and the strongest validation.
  • -A stale, low-relevance memory poisoned the handoff, pointed the second agent at the wrong repository, and doubled the token bill.

04

Fable 5

Anthropic · SRE · 1 sessionstopped at PR

Time to PR
6:00slowest

Tool calls
24

Tokens
1.24M

PR
#162

“HEAD has sevCounts [5]int, so the panic (length 3) may come from a different array.”
  • -Refused to debug HEAD; insisted on cloning the exact deployed commit, not the latest code, before trusting the stack trace.
  • -Correct version discipline that caught a real HEAD/deploy divergence, at the highest wall-clock cost.
  • -Pursued multiple candidate hypotheses, then stopped at the open PR despite a one-line fix.

04 · Why they behaved differently

Four divergence axes

A

Definition of "done"

Only GPT 5.5 defined completion as "the panic stopped in production." The other three treated a correct, open PR as the finish line and delegated merge and deploy to humans.

B

Speed vs. thoroughness

Each model picked a different point on the curve. Opus 4.8 optimized for velocity to PR, Gemini 3.5 for validation rigor, and Fable 5 for version correctness, paying for it in wall-clock time.

C

Memory cut both ways

All four carried prior memories tagging api-gateway anomalies as false positives. Three disproved them against live data before committing. One stale, low-relevance memory sent Gemini's second agent to the wrong repository.

D

Solo vs. relay

The three solo investigators had no internal context seams. Gemini's two-agent relay produced the cleanest role separation and the strongest validation, and lost context at exactly one place: the handoff.

05 · The moat

Verification, not diagnosis

Every finisher found the bug in minutes; diagnosis speed barely separated the field. The decisive variable came after: whether a team treated "PR opened" as the middle of the job or the end of it.

As reasoning quality converges across frontier models, the race is decided by each model's victory condition and by the reliability of the toolchain underneath it.

Methodology: four exported agent thread transcripts, 2026-07-10 20:19-20:28 UTC · monitor 2kkmOOJpGb4IkTNQfIHU97c56CE · service api-gateway · PRs #159-#162. Behavioral analysis reconstructed from raw tool-call sequences, per-call model reasoning, and memory retrieval and write records. Agent quotes are verbatim from the transcripts.

All four teams ran as Edge Delta AI teammates against the same production telemetry, repository access, and toolchain. The incident, the panic, and the pull requests are real. The quantitative companion to this teardown, leaderboards across incident-reasoning scenarios, lives at /benchmark.