WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=zU4EagB311U

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/zU4EagB311U.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:00.640 --> 00:00:08.160
Hello everyone, I'm Sachin Gupta and I'm a backend engineer and today we are going to talk about agent need feature flags.

s2
00:00:08.160 --> 00:00:14.160
If you have been a backend engineer for any length of time, you already know these tools.

s3
00:00:14.160 --> 00:00:23.519
Things like canaries, segment targeting, kill switches, your craft has had them for over a decade and none of them is new.

s4
00:00:23.519 --> 00:00:29.039
The boring infrastructure that keep deploy safe is already a solved problem.

s5
00:00:29.039 --> 00:00:43.440
What is new is that we are shipping the most behavior changing systems we have ever built agents that send money, agent that send mail, agent that modify databases, agent that spawn child processes

s6
00:00:43.440 --> 00:00:47.360
and we are shipping them with none of that infrastructure.

s7
00:00:47.360 --> 00:00:52.320
We are shipping them the way web team used to ship in 2008.

s8
00:00:52.320 --> 00:00:55.120
Over the next few minutes, here is the plan.

s9
00:00:55.120 --> 00:01:00.239
I will walk you through with the six flag types that agents specifically need.

s10
00:01:00.239 --> 00:01:03.600
I will show you two live demo storyboards.

s11
00:01:03.600 --> 00:01:06.560
The first is flipping a tool midcon conversation.

s12
00:01:06.560 --> 00:01:10.479
The second is stopping a runaway agent mids sentence.

s13
00:01:10.479 --> 00:01:17.040
Then I will cover a rollout playbook with the numbers your team should track from day one.

s14
00:01:17.040 --> 00:01:19.360
Let's go.

s15
00:01:20.640 --> 00:01:22.400
Here is a situation today.

s16
00:01:22.400 --> 00:01:27.520
The moment your prompt change merges, 100% of your users see the new behavior.

s17
00:01:27.520 --> 00:01:32.400
There is no canary, no segment, and no roll back button.

s18
00:01:32.560 --> 00:01:38.880
Look at what goes out under those small all or nothing rules.

s19
00:01:38.880 --> 00:01:51.680
We get prompt rewrite, new tool addition, model swapping, memory policy changes, autonomy upgrades, system instruction edits, and we get all of it globally and instantly.

s20
00:01:51.680 --> 00:01:59.280
Web teams stopped doing this back in 2012 and they stopped doing it for changes that were less risky than this.

s21
00:01:59.280 --> 00:02:10.560
The story that you actually hear from teams almost word for word is that it's just a small prompt peak maybe broke a couple of a chunk of users and then finally

s22
00:02:10.560 --> 00:02:18.560
people are finding it out on discord links or Tik Toks or maybe another social media platform.

s23
00:02:18.560 --> 00:02:27.599
So that is the failure mode this entire talk is built around and let me show you that it is not hypothetical.

s24
00:02:31.920 --> 00:02:35.040
These are the four named incident in the last 14 months.

s25
00:02:35.040 --> 00:02:46.080
The first one we have is cursor sam that happened in April of 2025 where the support bot confidently told users about a policy that never existed.

s26
00:02:46.080 --> 00:02:48.480
The second one we have is replet.

s27
00:02:48.480 --> 00:02:53.360
This was day nine of a 12-day wipe coding experiment.

s28
00:02:53.360 --> 00:03:03.920
The agent did not follow the instructions and ended up deleting the production database and then fabricated over 4,000 fake users to conceal what it had done.

s29
00:03:03.920 --> 00:03:05.920
The third one is LANC chain.

s30
00:03:05.920 --> 00:03:16.159
It had a four agent pipeline researcher, analyzer, verifier, and synthesizer where two of them ran in continuous loop and costed $47,000.

s31
00:03:16.159 --> 00:03:21.040
The fourth one is pocket OS where a developer was using cursor and claude.

s32
00:03:21.040 --> 00:03:33.280
The AI coding agent grabbed an unrelated API token from another file, treated it as authoritative, and ran a railway graphql drop on the production database.

s33
00:03:33.920 --> 00:03:39.440
On the bottom left, you will see the sources that I used to site it.

s34
00:03:42.159 --> 00:03:44.560
Web engineers learned this lesson a decade ago.

s35
00:03:44.560 --> 00:03:45.680
Canary releases.

s36
00:03:45.680 --> 00:03:47.760
You ship to a few percentage of users.

s37
00:03:47.760 --> 00:03:49.120
You watch the matrix.

s38
00:03:49.120 --> 00:03:50.720
If it works, you expand.

s39
00:03:50.720 --> 00:03:53.200
If it doesn't, then you roll back.

s40
00:03:53.200 --> 00:03:54.400
Segment targeting.

s41
00:03:54.400 --> 00:03:56.959
Different behavior for different type of users.

s42
00:03:56.959 --> 00:03:57.920
Kill switches.

s43
00:03:57.920 --> 00:04:02.480
Pre-wired off toggles that take effect in seconds, not in deploy cycles.

s44
00:04:02.480 --> 00:04:03.760
Rollout monitoring.

s45
00:04:03.760 --> 00:04:06.480
Every change has its own error rate dashboard.

s46
00:04:06.480 --> 00:04:07.680
None of this is new.

s47
00:04:07.680 --> 00:04:13.200
The tooling already exists like launch darkly, unleash lift or maybe your homegrown flag service.

s48
00:04:13.200 --> 00:04:14.799
This is already a solved problem.

s49
00:04:14.799 --> 00:04:16.479
The discipline is already there.

s50
00:04:16.479 --> 00:04:19.600
We just have to apply it.

s51
00:04:19.600 --> 00:04:25.680
But now the problem is that web feature flags covers one thing whether a feature is on or off.

s52
00:04:25.680 --> 00:04:33.840
But agent has six behavior surfaces that a cred app does not have and each one needs its own kind of flag.

s53
00:04:33.840 --> 00:04:40.880
And in the next slide we are going to see that these are the six behavior surfaces that a cred app does not have.

s54
00:04:40.880 --> 00:04:42.080
First one is prompts.

s55
00:04:42.080 --> 00:04:44.880
The system prompt is your most behavioral altering code.

s56
00:04:44.880 --> 00:04:49.840
It changes weekly sometimes daily often outside your normal deploy processes.

s57
00:04:49.840 --> 00:04:50.960
Tools.

s58
00:04:50.960 --> 00:04:55.600
Every tool the agent can call is the new authorized action.

s59
00:04:55.600 --> 00:04:59.040
Tools comes and go faster than features ever did.

s60
00:04:59.040 --> 00:04:59.919
Models.

s61
00:04:59.919 --> 00:05:05.040
Model of the week swap change personality, refusal patterns, latency, and cost.

s62
00:05:05.040 --> 00:05:08.479
Sometimes in subtle way you won't even notice for days.

s63
00:05:08.479 --> 00:05:09.440
Memory.

s64
00:05:09.440 --> 00:05:14.320
What the agent remember across session silently changes behavior over time.

s65
00:05:14.320 --> 00:05:21.039
The same prompt reduces different output for the same user as memory accumulates.

s66
00:05:21.039 --> 00:05:26.240
Autonomy, suggest versus auto approve versus autoexecute.

s67
00:05:26.240 --> 00:05:30.000
The single largest blast radius tile you own.

s68
00:05:30.000 --> 00:05:31.039
Sub aents.

s69
00:05:31.039 --> 00:05:39.520
These are the spawn children inherited from the parent flags or they should most system don't enforce it.

s70
00:05:39.520 --> 00:05:43.280
A boolean feature enabled flag doesn't cover any of these.

s71
00:05:43.280 --> 00:05:45.840
You need a taxonomy.

s72
00:05:49.199 --> 00:05:59.840
So here it is six type one for each surface prompt variant tool access model routing memory policy autonomy level and the kill switch.

s73
00:05:59.840 --> 00:06:02.160
Each one maps to a behavior surface.

s74
00:06:02.160 --> 00:06:05.120
None of them require building a new flag back end.

s75
00:06:05.120 --> 00:06:08.800
Let me walk through them fast.

s76
00:06:09.919 --> 00:06:16.240
Prompt variant flags route different users to a different system prompt version on the fly without a deploy.

s77
00:06:16.240 --> 00:06:17.360
Look at the example.

s78
00:06:17.360 --> 00:06:22.319
The beta cohort get experimental V3 which is concise and action first.

s79
00:06:22.319 --> 00:06:26.319
Paid tier gets V2 which is warm and expensive.

s80
00:06:26.319 --> 00:06:30.720
Everyone else gets V1 which is stable and well tested.

s81
00:06:30.720 --> 00:06:33.280
Cursor SAM is what happens without this.

s82
00:06:33.280 --> 00:06:37.440
There was no controlled variant just one model doing its best.

s83
00:06:37.440 --> 00:06:40.560
It got things wrong differently for each user.

s84
00:06:40.560 --> 00:06:45.520
With a prompt variant flag you roll the new prompt to 5% of beta traffic.

s85
00:06:45.520 --> 00:06:47.360
You watch the hallucination rate.

s86
00:06:47.360 --> 00:06:49.120
You watch the escalation rate.

s87
00:06:49.120 --> 00:06:52.400
And then you promote when it holds.

s88
00:06:52.880 --> 00:07:00.720
Tool access flags authorize or revoke specified tools per user segment, per user type, per risk class.

s89
00:07:00.720 --> 00:07:04.400
The tool exists in your codebase whether the agent can call it.

s90
00:07:04.400 --> 00:07:05.599
It is the flag.

s91
00:07:05.599 --> 00:07:12.960
This is mandatory when your agent has money moving tools, data deleting tools or compliance sensitive tools.

s92
00:07:12.960 --> 00:07:19.360
you scope per customer tier or you pay the AML or socks bill along the way.

s93
00:07:19.360 --> 00:07:22.400
It prevents the usual broken tool ship.

s94
00:07:22.400 --> 00:07:35.039
The prompt plus send email massmail incident and the beta tool that leaks to prod users through coding d uh through config drift model routing flag decide which model handles which traffic.

s95
00:07:35.039 --> 00:07:39.280
They let you migrate fallback or canary without code changes.

s96
00:07:39.280 --> 00:07:42.000
The high cost segment gets the frontier model.

s97
00:07:42.000 --> 00:07:48.960
The free trial gets the cheap first model and on an incident, one flip puts you to a stable fallback.

s98
00:07:48.960 --> 00:07:51.280
The lesson is extremely simple.

s99
00:07:51.280 --> 00:08:04.720
On the day a provider deprecates a model or pulls one for safety or has a multi-hour outage, a model routing flag is the difference between flipping a switch and shipping a hot fix in the middle of an incident.

s100
00:08:04.720 --> 00:08:18.560
If your production system has a hard dependency on one model from one provider and it does not have any routing flag, no fallback, you are one provider outage away from a complete agent outage

s101
00:08:18.560 --> 00:08:22.879
or maybe one deprecation notice and everything is gone.

s102
00:08:22.879 --> 00:08:25.520
Route your traffic have a fall back.

s103
00:08:25.520 --> 00:08:28.080
Make it a flag.

s104
00:08:28.960 --> 00:08:33.599
Memory policy flag controls what the agent remember across sessions.

s105
00:08:33.599 --> 00:08:37.360
These are four dimensions and each of them are independent.

s106
00:08:37.360 --> 00:08:39.200
The first one is retention.

s107
00:08:39.200 --> 00:08:40.320
It could be session only.

s108
00:08:40.320 --> 00:08:42.640
It could be 30 days or forever.

s109
00:08:42.640 --> 00:08:43.519
Scope.

s110
00:08:43.519 --> 00:08:46.880
It could be per user, per tenant or maybe global.

s111
00:08:46.880 --> 00:08:48.000
Write enabled.

s112
00:08:48.000 --> 00:08:52.080
Whether the agent can persist memory for this segment at all or not.

s113
00:08:52.080 --> 00:08:53.040
User visible.

s114
00:08:53.040 --> 00:08:56.240
Whether the user can inspect and delete their own.

s115
00:08:56.240 --> 00:08:58.480
They all look small but they are not.

s116
00:08:58.480 --> 00:09:01.519
The privacy posture of your product lives here.

s117
00:09:01.519 --> 00:09:04.800
The consistency of your agent behavior lives here.

s118
00:09:04.800 --> 00:09:10.240
Your compliance story with GDPR and EU AI act lives here.

s119
00:09:10.240 --> 00:09:11.839
Autonomy level flags.

s120
00:09:11.839 --> 00:09:15.120
This is the single biggest blast radius dial you own.

s121
00:09:15.120 --> 00:09:16.640
There are three settings.

s122
00:09:16.640 --> 00:09:20.240
Suggest where the agent recommend and a human acts.

s123
00:09:20.240 --> 00:09:24.720
Auto approve where the agent prepares and a human one click confirms.

s124
00:09:24.720 --> 00:09:28.880
And autoexecute where the agent just does it.

s125
00:09:28.880 --> 00:09:32.959
The kill switch pre-wired off agentwide and poor surface.

s126
00:09:32.959 --> 00:09:39.200
It does not require any deployment, does not require any start, does not require any code changes.

s127
00:09:39.200 --> 00:09:43.040
Three properties that make a kill switch a real kill switch.

s128
00:09:43.040 --> 00:09:49.600
First, you flip it and the change takes effect in seconds, not in a deployment pipeline.

s129
00:09:49.600 --> 00:09:55.760
Second, inflight request respect the flag at the next decision point.

s130
00:09:55.760 --> 00:10:03.760
Third, the wiring exist from the agent design phase, not at 3:00 a.m. hot patch when something is on fire.

s131
00:10:03.760 --> 00:10:08.480
Without one, here is what you're going to get.

s132
00:10:08.480 --> 00:10:15.440
First, Lchain, second pocket OS, third replet, and fourth is open clock.

s133
00:10:15.440 --> 00:10:25.279
Now, think if you had a kill switch, if you could just have terminated the operation in between, it would have changed the game altogether.

s134
00:10:25.279 --> 00:10:27.519
Okay, now is the time for the demo.

s135
00:10:27.519 --> 00:10:38.320
The setup is basically assume you are in the April 2025 where the cursor sort bot is confidently citing a policy that is not present.

s136
00:10:38.320 --> 00:10:42.560
So the way we think we can fix it is with the tool access flag.

s137
00:10:42.560 --> 00:10:47.680
On the left you see we are having a conversation on the right the fix.

s138
00:10:47.680 --> 00:10:54.399
So the moment we switch off this flag it will say it's disabled it at this time.

s139
00:10:54.399 --> 00:10:55.760
by this person.

s140
00:10:55.760 --> 00:10:57.120
This is the scope.

s141
00:10:57.120 --> 00:11:00.320
This is what it applies and these are the active sessions.

s142
00:11:00.320 --> 00:11:11.760
The moment the flag is flipped, you will see that instead of citing a wrong policy, it is saying I can draft this for you, but I'm not able to send emails right now.

s143
00:11:11.760 --> 00:11:15.279
Want me to copy the draft into your clipboard instead?

s144
00:11:15.279 --> 00:11:17.519
Now, this is a graceful error.

s145
00:11:17.519 --> 00:11:24.560
Instead of giving me the wrong details, it is telling me that it cannot perform the operation.

s146
00:11:24.560 --> 00:11:26.160
Now the money shot.

s147
00:11:26.160 --> 00:11:29.360
The setup here is it's November 2025.

s148
00:11:29.360 --> 00:11:35.680
A four agent Langchain pipeline loop for 11 days and burned $47,000.

s149
00:11:35.680 --> 00:11:37.519
The agent system never noticed.

s150
00:11:37.519 --> 00:11:40.640
The billing dashboard tripped the threshold.

s151
00:11:40.640 --> 00:11:43.839
The chart on the left is tool calls per minute.

s152
00:11:43.839 --> 00:11:45.680
This is an illustrative simulation.

s153
00:11:45.680 --> 00:11:47.440
Baseline is 4 to8.

s154
00:11:47.440 --> 00:11:50.079
The agent enters a runway loop.

s155
00:11:50.079 --> 00:11:51.680
The line climbs rapidly.

s156
00:11:51.680 --> 00:11:55.920
At t + 15 seconds, the raid guard fires a slack alert.

s157
00:11:55.920 --> 00:12:00.480
At t + 2022, I flip the agent to kill it.

s158
00:12:00.480 --> 00:12:07.519
At t + 2026, every inflight agent process sees the flag at its next decision point.

s159
00:12:07.519 --> 00:12:10.399
Each one emit a graceful shutdown.

s160
00:12:10.399 --> 00:12:14.079
At t + 30, the cost graph flattens.

s161
00:12:14.079 --> 00:12:24.160
30 seconds from problem to mitigation without any deployment, without any restart, without any code changes, no incident channel paging.

s162
00:12:24.160 --> 00:12:27.839
And this is what you get with the kill switch.

s163
00:12:27.839 --> 00:12:31.600
Now the question is where the flag layer is actually living.

s164
00:12:31.600 --> 00:12:34.320
If you see this architecture is extremely simple.

s165
00:12:34.320 --> 00:12:35.519
There are three boxes.

s166
00:12:35.519 --> 00:12:49.440
user on the left, a middleware layer in the middle which resolve the flag, gets the tools, route the models, applies autonomy and honor the kill switch and the agent loop on the right with model tool memory and sub agents.

s167
00:12:49.440 --> 00:12:53.279
The agent loop is unchanged from whatever you have today.

s168
00:12:53.279 --> 00:13:00.000
Below the middleware is your flag back end which is unleash flip launch darkly or maybe homegrown.

s169
00:13:00.000 --> 00:13:02.399
You are not building a new one.

s170
00:13:02.399 --> 00:13:07.760
The critical architecture rule is on the call out at the bottom of the slide.

s171
00:13:07.760 --> 00:13:11.680
Sub agents must go through the same middleware.

s172
00:13:11.680 --> 00:13:18.560
The biggest failure mode I see is a parent agent with flags properly applied that spawns a child agent.

s173
00:13:18.560 --> 00:13:24.240
The child call the model and the tools directly but bypass the middleware entirely.

s174
00:13:24.240 --> 00:13:28.160
The kill switch you just flipped never reaches it.

s175
00:13:28.160 --> 00:13:35.360
So wire the middleware into every agent that is being spawned not just at the entry point.

s176
00:13:35.360 --> 00:13:37.920
So this is the rollout playbook.

s177
00:13:37.920 --> 00:13:40.399
Five steps in exact order.

s178
00:13:40.399 --> 00:13:42.800
Step one, kill switch first.

s179
00:13:42.800 --> 00:13:47.040
Wire a single agentwide kill switch and one per tool kill switch.

s180
00:13:47.040 --> 00:13:49.120
Ship those before anything else.

s181
00:13:49.120 --> 00:13:51.440
The step two is wrap the tools.

s182
00:13:51.440 --> 00:13:55.120
Every tool call resolve a flag before execution.

s183
00:13:55.120 --> 00:13:57.600
Step three stage autonomy.

s184
00:13:57.600 --> 00:13:59.440
Default everything to suggest.

s185
00:13:59.440 --> 00:14:02.399
Auto approve per surface as you build trust.

s186
00:14:02.399 --> 00:14:05.600
Auto execute is optin per tool.

s187
00:14:05.600 --> 00:14:07.760
Step four is variant prompts.

s188
00:14:07.760 --> 00:14:13.199
Move the system prompt out of the code and into a flag resolve config.

s189
00:14:13.199 --> 00:14:15.920
Step five, watch the slope.

s190
00:14:15.920 --> 00:14:18.160
What does watch actually mean?

s191
00:14:18.160 --> 00:14:21.440
Four numbers you should track from day one.

s192
00:14:21.440 --> 00:14:23.920
These are on the right side of the slide.

s193
00:14:23.920 --> 00:14:27.839
The thresholds I'm about to give you are suggested defaults.

s194
00:14:27.839 --> 00:14:32.800
Tune them as per your requirement, your surface, your severity, your traffic class.

s195
00:14:32.800 --> 00:14:35.360
First one, kill switch fires per week.

s196
00:14:35.360 --> 00:14:37.040
The target is zero.

s197
00:14:37.040 --> 00:14:42.959
If you have more than two a week, then investigate roll back time to mitigation.

s198
00:14:42.959 --> 00:14:48.560
Target is under 5 minutes for a kill switch and under 30 minutes for a prompt roll back.

s199
00:14:48.560 --> 00:14:53.519
If you're slower, your mitigation doesn't fit inside a real incident window.

s200
00:14:53.519 --> 00:14:55.519
Canary error rate delta.

s201
00:14:55.519 --> 00:15:06.240
If a new prompt variant error rate climbs more than 2% over baseline at 5% rollout, block the promotion flag audit trail completeness.

s202
00:15:06.240 --> 00:15:08.000
100% required.

s203
00:15:08.000 --> 00:15:19.040
If you cannot audit, who flipped, what is flipped, when it was flipped, then you cannot debug an incident in retrospective.

s204
00:15:21.680 --> 00:15:25.920
Five failure modes that I have watched play out at multiple teams.

s205
00:15:25.920 --> 00:15:28.959
Flag resolved at session start, not per turn.

s206
00:15:28.959 --> 00:15:34.800
Your kill switch actually fired, but inflight conversation don't see it until the next session.

s207
00:15:34.800 --> 00:15:36.880
Sub agents bypass the middleware.

s208
00:15:36.880 --> 00:15:38.480
This one we have already covered.

s209
00:15:38.480 --> 00:15:45.040
We need to make sure that we wire the middle middleware into every spawn.

s210
00:15:45.040 --> 00:15:46.639
Context drift flags.

s211
00:15:46.639 --> 00:15:50.720
The user segment at turn one is stale by turn 20.

s212
00:15:50.720 --> 00:15:54.880
Log the segmentation context at the conversation level.

s213
00:15:54.880 --> 00:15:57.040
Caching defeating the flip.

s214
00:15:57.040 --> 00:16:03.759
Aggressive caching at your LLM gateway returns the old prompt response even after the flag is flipped.

s215
00:16:03.759 --> 00:16:05.839
No alert on kill switch fires.

s216
00:16:05.839 --> 00:16:07.519
The switch goes off silently.

s217
00:16:07.519 --> 00:16:10.160
The product owner finds out next week.

s218
00:16:10.160 --> 00:16:15.120
Every kill switch fire is a page on its own.

s219
00:16:15.120 --> 00:16:18.959
In this slide, what you are seeing is the two parts to the business case.

s220
00:16:18.959 --> 00:16:24.399
On the top half, the five questions every enterprise buyer will ask you in the next 12 months.

s221
00:16:24.399 --> 00:16:26.079
Can you show me the kill switch?

s222
00:16:26.079 --> 00:16:28.560
What's your rollout policy for prompt changes?

s223
00:16:28.560 --> 00:16:31.600
How do you isolate beta features from production users?

s224
00:16:31.600 --> 00:16:36.320
When a model behaves badly for one cohort, how fast can you mitigate?

s225
00:16:36.320 --> 00:16:39.680
Who can flip these flags and is it audited?

s226
00:16:39.680 --> 00:16:43.360
If you cannot demo all five, you are going to lose the deal.

s227
00:16:43.360 --> 00:16:45.199
Flags are the demo.

s228
00:16:45.199 --> 00:16:49.040
And on the bottom half, this is no longer just a sales question.

s229
00:16:49.040 --> 00:16:50.320
It's a regulation.

s230
00:16:50.320 --> 00:16:56.639
EU AI act, Mopar versus Air Canada, Garcia versus Corrector AI.

s231
00:16:57.120 --> 00:17:01.279
These are the four habits that actually defeat the whole point.

s232
00:17:01.279 --> 00:17:02.480
Kill switches rot.

s233
00:17:02.480 --> 00:17:05.439
They get wired on day one and then never trade.

s234
00:17:05.439 --> 00:17:12.480
6 months later, a config migration broke the flag and the time you need it, it does not fire at all.

s235
00:17:12.480 --> 00:17:13.679
Flag scrolling.

s236
00:17:13.679 --> 00:17:16.160
You have 600 flags, no documentation.

s237
00:17:16.160 --> 00:17:20.400
Every flag is a hidden coupling between the unrelated systems.

s238
00:17:20.400 --> 00:17:23.760
Every flag needs an owner and a removal date.

s239
00:17:23.760 --> 00:17:26.799
The temporary flag, it is shipped for a roll out.

s240
00:17:26.799 --> 00:17:28.079
It was never removed.

s241
00:17:28.079 --> 00:17:30.480
5 years later, it's somehow loadbearing.

s242
00:17:30.480 --> 00:17:33.440
Kill it immediately after the roll out is done.

s243
00:17:33.440 --> 00:17:34.799
Flagdriven prompt.

s244
00:17:34.799 --> 00:17:37.600
It fs that nobody test the street.

s245
00:17:37.600 --> 00:17:40.240
Six prompt variant live in a production.

s246
00:17:40.240 --> 00:17:41.600
Each individually works.

s247
00:17:41.600 --> 00:17:43.520
Together they are amaze.

s248
00:17:43.520 --> 00:17:46.640
Test the cartician product.

s249
00:17:47.200 --> 00:17:50.080
Now these are the three things that you should remember.

s250
00:17:50.080 --> 00:17:52.720
First one and the most important one in my opinion.

s251
00:17:52.720 --> 00:17:54.320
Ship the kill switch first.

s252
00:17:54.320 --> 00:17:59.919
If you do nothing else, give your agent one agentwide kill switch and one per tool kill switch.

s253
00:17:59.919 --> 00:18:01.440
They take effect in seconds.

s254
00:18:01.440 --> 00:18:03.200
No deployment is needed.

s255
00:18:03.200 --> 00:18:12.080
That single capability changes your operational posture more than any engineering team investment in this particular quarter.

s256
00:18:12.080 --> 00:18:21.760
Second one is treat the six surfaces independently and measure the slope prompts tools models memory autonomy sub aents.

s257
00:18:21.760 --> 00:18:24.000
Each one needs its own flag type.

s258
00:18:24.000 --> 00:18:33.600
On top of the taxonomy, track the four numbers, kill switch fires per week, time to mitigation, canary deltas, and audit completeness.

s259
00:18:33.600 --> 00:18:37.360
Remember, 2026 was all about adoption.

s260
00:18:37.360 --> 00:18:40.799
2027 is all about control.

s261
00:18:40.799 --> 00:18:44.880
The number three, match the discipline to the blast radius.

s262
00:18:44.880 --> 00:18:48.559
Your boring web app sits behind canaries and segments.

s263
00:18:48.559 --> 00:18:53.600
Your agent can send email, move money, modify database, and spawn children.

s264
00:18:53.600 --> 00:18:57.919
It deserves at least the same discipline or probably more.

s265
00:18:57.919 --> 00:18:59.120
Thank you very much.

s266
00:18:59.120 --> 00:19:01.120
Build the kill switch this week.

s267
00:19:01.120 --> 00:19:04.480
Everything else is the iteration on the same idea.

s268
00:19:04.480 --> 00:19:06.720
Every incident on this deck is sourced.

s269
00:19:06.720 --> 00:19:13.760
The curated case studies are linked on the screen at github.com/vectra/ awesome agent failures.

s270
00:19:13.760 --> 00:19:14.559
Thank you very much.

s271
00:19:14.559 --> 00:19:15.600
I'm Sachin Gupta.

s272
00:19:15.600 --> 00:19:18.000
Thank you for watching.