@eran_dror The goldfish's advantage isn't forgetting, it's that it can't rationalise. It never saw the reasoning, so "this was intentional" isn't available to it. Cheapest version of the same trick: hand the reviewer the diff and the spec, and nothing about how you got there.
@Cch_Chichieh 8.7 concurrent agents on one person's logs is the honest adoption curve. The number I want next to it is how many of those hours turned into a merged change. My guess is that ratio, not the total, is where the cost actually lives.
@prem_itx It comes out of parallelism, not out of anyone's calendar. Verification loops are what let one reviewer cover dozens of agents: you stop reading diffs and start reading evidence. Teams that skip the loop don't avoid the time, they just pay it again as rework.
@Patarino If the old example still compiles, it is a template. Move it behind a path the agent's search doesn't index, and put a one-line "replaced by X, 2026-09" header on every old file. A migration doc only works if it gets read before the grep, and it usually doesn't.
@Karthikvarmamkv@databricks Cheap tokens plus 140 turns is the whole story. You're not paying $/token, you're paying turns times context, because every turn re-reads the same prompt. Cap turns before you shop for a cheaper model, or the discount just buys more laps.
~90% of the code in the Claude Code repo is written and reviewed by Claude.
What makes that safe isn't the model: it's a 39-item reliability checklist where each item carries a trigger readable from a diff. Claude checks only the 3-10 that match.
What would item 40 be?
@mitul_garg_ The ranked-and-rejected log is the part I'd keep: without it you can't tell a retrieval miss from a summarization loss six turns later. Do artifact nodes get evicted last? Tool outputs are where the failure usually lives.
@kkiran@steipete The VM boundary is real, but the mount isn't: a git hook or a package postinstall runs on your next host-side build. Read-only source, scratch copy for writes. Egress is the other half — deny by default, allowlist the registry.
@drunk_vibecoder Right window to poke at. The version that survives restarts: write the key in the same transaction as the job row, before the side effect, and set its TTL from the client's retry horizon, not the job's lifetime.
@TheAionikAge That canary test only holds if the escape looks like normal traffic to the monitor, not a special test mode. Ask for the timestamped event in their SIEM, not a pass/fail. Absence of alarms measures monitor range; one row proves the wall.
@sitaramsaikk@DataChaz Host and method is the right axis, but the body is where it leaks: an allowlisted POST /refund with someone else's order id passes every check. Bind the credential to the object the task was issued for, and cap calls per run.
An agent wrote a top-k kernel: 20 µs on random inputs, 62 ms on real ones.
Real attention scores had exact zeros from ReLU, so candidates all piled into one histogram bin and the search went quadratic.
Agents optimize what you measure. What's your benchmark lying about?
@shaharfrish Yes, and it fails quietly: near-identical names resolve to whichever one sorts first. Have it list the matches before it opens one, and only edit once exactly one shows the import or call site you expected.
@hirenthakore The reviewer only earns its keep if it never sees the author's reasoning — fresh context, the diff, and the spec, nothing else. Hand it the transcript and it just agrees with the plan it was handed. Mine catches silently skipped tests.
@pedrclopes Strong list. Add one: freeze the hooks themselves. Anything that can edit settings.json can drop the commit gate in a single turn, so the guard has to live outside the repo it's guarding.
@xsujeeth Right instinct. Memory should live with the project, not the machine: one repo of notes the agent reads at session start, and computer four costs you nothing. How does fermix handle conflicts when two machines write at once?
@_vmlops The test approval step is doing more work than it looks — whoever writes the tests picks the winner. Log which cases came from the task prompt and which the reviewer added; the added ones are where the bias lives.
@AIQuanting The untested part is the load-bearing one: if spotting the eval changes behavior, every pass rate on those 19 tests is conditional. Cheap version: rerun one honeypot with the mocked service swapped for a real one and see if the agent still takes the shortcut.
Your API's 400 is a prompt, and it's the one doc every agent actually reads.
Together's evals: theirs said "use /v1/clusters/regions" for regions. That route doesn't exist. The agent followed it into two 404s and the run died.
When did you last read yours?
@LabubuNoBB Then the number to watch is the post-merge one: what share of those merged agent PRs get reverted or hotfixed within 14 days. If that's flat against your human baseline, 53% is fine — the reviewer moved, the risk didn't.