Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati

AI Engineer · 21 min · 169 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:12Hi everyone, I'm Arena, former engineer at Microsoft and Supercell.
00:18And today I want to talk about auto research in a multi- aent AI village.
00:24I will use a video game like AI Village as a running example here, but the broader question is one I think many AI engineers are starting to run into.
00:34How do we evaluate and improve agents that carry state over a long period of time?
00:43Before I get into the auto research layer, I want to talk a bit about project paradox.
00:49We developed project paradox at supercell's AI innovation lab.
00:54Me and my teammate are natural money.
00:59We built a modular AI framework that allows any developer to plug in intelligent autonomous agents within a video game that can interact, compete or cooperate with other players or agents as well
01:14and place them uh and make them into dynamic game companions.
01:20Now to give examples of what these agents can do, the agents can move with intent.
01:26They can go to any location or person and they're guided by their own memories, emotion, or curiosity.
01:35These agents can interact with the world.
01:38They can pick up objects, drop them anywhere, and they're also aware about the context in their own environment, such as the objects or other characters or agents as well.
01:49I would also like to note that game developers can also add new actions for these agents to accomplish within our framework as well.
01:58Instead of just dropping or uh placing objects, agents can also obviously react to what's happening around them.
02:08And these events that happen around them affect their own beliefs and emotions on the fly as well.
02:16And of course, it wouldn't be complete if agents can't start conversations, right?
02:21agents can in this scenario approach other agents or even the player as well and this makes the game feel more alive.
02:31And of course these conversations are stored within their memory and is according to their own um and affect their own emotions and beliefs or goals as well.
02:43And al together these agents make our multi- aentic framework.
02:51Um yeah yeah one second so the architecture was intentionally stateful behind this.
03:02The first important part was per agent memory.
03:07Each agent has its own memory namespace backed by rag.
03:12So memory did not bleed between agents.
03:16Second, we tracked emotion as a small vector.
03:19So after an event or conversation, the system could update values like joy, sadness, fear, anger, or disgust.
03:29Third, agents had belief scores towards other agents and the player.
03:36You can think of this as a trust matrix.
03:38Basically like after the interaction happens, the LM basically decides whether the trust score should go up, down, or whether it shouldn't change at all.
03:48And fourth, every memory receives an important score.
03:53Um to to explain this better, like let's say you had dinner a few days ago, you probably wouldn't remember what you had for dinner, right?
04:01But um if someone was murdered a few days ago, you definitely remember that.
04:08So the agent will evaluate or the LM will evaluate uh an important score of an event and if it crosses a threshold, it will store that specific memory uh in a separate cache
04:21so that important context can be retrieved better later on.
04:28And here's an example of it just working.
04:32Um, we going to ask one of the characters to go on a picnic with us.
04:36Here, uh, our character Blossom um, decides to pick up a pastry and go to the picnic area because we asked her to do so.
04:46Keep in mind during the conversation in the background, she plans all of these sequence of actions to accomplish.
04:54And one when we talk to her afterwards, she will also reply within context as well.
05:03Yeah.
05:06But this is where an interesting problem actually started.
05:11As you saw in the last example, like for shortterm game play, this our architecture worked pretty well.
05:19like a character could make a plan, move around, talk and remember the recent interaction and respond to us or other characters as well.
05:28But over longer horizons, this is where we notice the social consistency start to get weaker.
05:36So in this example, we have one agent spreading a rumor about a sale on mangoes to another agent and that agent receives that information and goes and tells another agent about it.
05:49Later on, after a number of events that occurred in between, when the player asks one of the agents about the mangoes, it doesn't exactly store that context that we were expecting or it doesn't give us the context that we kind of wanted to.
06:06And this is where things are starting to get messy naturally.
06:11Like the system may remember the rough topic but lose the source of the topic.
06:16A rumor may become certain instead of just a rumor like the agent might state it as a fact or um an agent might know a fact but fail to execute fail to remember it while creating a plan for its actions.
06:32So the question here became how do we improve a multi-agentic system over longunning social behavior and not just over one response.
06:44And this is where we wanted to bring in auto research.
06:48As you all know, uh few months ago, Karpathi posted out auto research and this this made us immediately very curious.
06:58Uh perhaps we can make the system run experiments uh on itself and can we use this for our system as well.
07:07So what we understood is instead of manually tuning a prompt or watching one nice demo, we could define sen a scenario suit, run the agents, collect traces, score the behavior
07:21and change a small policy surface and only keep the changes that actually improve the score.
07:28And this is where we're trying to bridge project paradox with auto research.
07:33So at this point basically our multi-agentic framework project paradox is more like a lab bench and auto research becomes the experimental loop around it.
07:45And importantly this is not only about improving rag retrieval.
07:50The broader framing is optimizing the agent protocol like how do agents write memories, retrieve them, communicate uncertainty, update trust attribute sources and replan around new facts.
08:05basically um yeah in this context uh oh yeah in this context art research is not another agent in the village like I said it's a meta system outside the village
08:24the villagers have local perspectives of course they only know what they saw heard remembered or inferred because there isn't a common memory database in between them.
08:37Information only travels once uh other agents communicate them properly.
08:43The auto research layer has a different job here.
08:46It reads the full traces of a run, compares what happens against the scenario ground truth, uh scores the behavior and proposes a constrained change to the agent protocol or cognitive policy.
09:03Then it reruns the scenario and asks society level behavior like did society level behavior get better.
09:10This is the key shift we were trying to look for.
09:14So we were no longer evaluating one answer.
09:17We were evaluating an entire run.
09:21And this is what one of the loops would look like.
09:25Like first we define a control scenario which I'll elaborate a bit more about later.
09:31For example, one agent learns a public fact or one agent hears a rumor.
09:37Uh that could be a controlled scenario.
09:40Then we run the simulation.
09:42During the run, we collect structured traces, observations, conversations, memory rights, retrievals, belief updates, whatever is relevant to us in that case, we collect.
09:55Then we score this behavior.
09:58Did the information spread as we expected it to?
10:01Did the source attribution survive?
10:04Such as, does the agent remember who started the rumor?
10:08Did uncertainty stay uncertain?
10:10Did agents act on what they actually knew?
10:14And then the auto research layer here proposes a small policy change.
10:20And this is important.
10:22It should not rewrite the whole application of course.
10:26it should only edit a controlled policy surface and then we rerun.
10:31If the score improves and the guardrails hold, we we keep the improvement and if not we simply just revert back.
10:43And talking about controlled scenarios, the reason why uh scenario design matters is that social behavior is otherwise a bit fuzzy.
10:55uh in general in the sense if you just let the agents in our environment wander around it might look cool and you might get nice interactions but it's actually very hard to evaluate on whether the system actually improved.
11:12So this is why we believe you need controlled scenarios.
11:16For example, one scenario could test a public fact diffusion.
11:21Let's say agent A learns uh the bakery will close tomorrow.
11:26Do the right agents learn it?
11:28Do they remember who said what?
11:31Do they rem do they change their plans based on this fact?
11:35Another scenario could test rumor uncertainty.
11:39Agent let's say agent A hears that agent C might leave the village.
11:45When this rumor spreads, does might leave suddenly become is leaving or does it stay as might leave?
11:54Like does it become a fact or does it still stay as a a rumor?
11:59Another scenario could test replanning.
12:03The group has a plan, but one agent learns, let's say, the route they wanted to take is blocked.
12:10Do agents update this and communicate this uh with each other to avoid uh a improper plan or scale actions?
12:20The point is not that these exact scenarios are universal here.
12:24The point we're trying to make is that long horizon agent behavior needs scenario suits.
12:32And talking about our Mango example again, after running one of our auto research loops, this time after uh a a long pro period of time, when the player finally asked one of the agents about the sale on mangoes,
12:49we did find that u the the agent was able to respond within context this time like compared to last time.
13:02Um yeah and for this talk the form the exact formula we believe is less important than the shape of the scorecard.
13:13Uh you do not want a single vague met metric like agent quality.
13:20This will hide all the interesting failures.
13:22Instead you want a balanced scorecard.
13:26For diffusion, you might measure reach like how many agents know the fact after end steps.
13:33For provenence, you measure source retention among agents who know it.
13:40How many remember it, where it came from, etc. For rumors, you can measure uncerny preservation and false surn rate.
13:48For planning, you can measure action consistency and time to replan.
13:53And for privacy, you can measure containment.
13:57This matters because optimizing only one metric can create bad behavior.
14:02Because let's say if you only optimize for diffusion, the agents may learn to overshare everything.
14:10And let's say if you only optimize for memory recall, you might create noisy or steal um like memories.
14:17So this scorecard is what keeps the system honest and prevents the auto research agent from gamifying the system to just increase one specific score.
14:32The other important engineering lesson that we learned over this project is that um it's important to keep the editable surface really small.
14:44The auto research layer should not have permission to randomly rewrite the whole codebase.
14:51Instead, it's really important to freeze the harness, the scenarios, and the metrics.
14:58So, we're only exposing the part of the system that we actually want to optimize.
15:03Here in project paradox for us that meant things like memory writing policy, retrieval policy, communication prompt, belief, trust rules, source attribution, replanning triggers, etc. This gives the search pro process room to improve behavior, but also prevents it from gaming the evaluation directly as we mentioned before.
15:28And this is the difference between the LM writing random patches versus the LM actually searching within a controlled policy space.
15:40And here are here are examples of the kind of changes I want this kind of loop to search over.
15:48If if source attribution disappears, the policy change might be preserve source in memory and uh write uh memory rights and summaries.
15:59If rumors harden into facts, the policy change might be stored confidence mark firstand versus secondhand and require hedging when retelling uncertain claims.
16:10If if facts if public facts stay local, the policy change might be classify useful public facts differently and make agents proactively share important source evidence.
16:22The key is that these are small changes to the agent protocol, but they can have larger effects on a society level behavior for multi-agentic systems.
16:34This is also where I kind of want to be careful about our claims here because with we believe without repeated current loop results like I wouldn't say the system
16:46just generally improved.
16:49We're trying to say this is the right kind of surface to expose to an auto research layer uh loop because it is small enough to control but it's still rich enough to change the social behavior to some extent at least.
17:05And the biggest lesson for me perhaps was that memory is not enough here.
17:12You can add a rag memory to an agent and still not get the current long-term uh horizon behavior that you were looking for.
17:22Um because agents need to sometimes know where that information came for uh came from.
17:28You need to preserve whether it was firsthand, secondhand, verified or uncertain.
17:34Sometimes you need to separate raw episodic memories from what the agent currently believes too.
17:39And you need to test behavior through scenarios not not just through vibes.
17:44So the other lesson is that uh roll back also is not optional.
17:50When you optimize social behavior, a change can improve one thing and damage another.
17:56So, a policy that spreads public facts uh faster might also leak private information.
18:02A policy that increases recall might increase stale memory usage.
18:07So, the loop should basically be like a ratchet.
18:10Try a change, score it, keep it only if the scorecard improves and guard rails whole.
18:19And we we definitely believe this is not only relevant for game agents cuz although I gave you an example using a game village um we believe like let's say for example support agents support agents need to know which policy update comes from where right and whether it supersedes an older answer.
18:40Personal assistants for example need to remember commitments that they previously made and h make corrections if uh if the user u wants to change those personal commitments.
18:53Research agents need pro uh provenence citations contradiction handling and hypothesis updates.
19:00Coding agents need longunning context across issues, files, teammates and changing requirements.
19:07Workflow agents need access controls, handoffs, and replplanning when the world changes.
19:13All of these systems have the same underlying problem.
19:17They maintain state over time.
19:20And that state affect affects future action.
19:24So they need control scenarios and behavioral scorecards is what we are proposing.
19:31So again in brief, a recipe for long horizon agents.
19:36If there's one practical recipe I want you to take away, freeze the harness, define scenarios, log traces, score behavior, and expose only a small policy surface.
19:51Search over these changes, keep only changes that survive your measurement.
19:56And this is an engineering pattern that we believe would uh make sense for longunning agents.
20:03The real question we believe is across controlled runs, does the system behave better?
20:13To close, project paradox start as an attempt to make game agents feel alive in a 3D world.
20:19But the deeper engineing problem was not animation or dialogue for us.
20:24It was the state such as which agent knows what, which agent told whom, what is true, uncertain or outdated, and do agents act on what they remember.
20:36Otter research.
20:37Otter research gave us a way to approach this a bit more systematically.
20:42Not by trusting one demo and not by endlessly handtuning prompts, but by running control experiments and keeping only the changes that survived our measurement.
20:53Long horizon agents need experiments and not just prompts.
20:57And I hope that's the takeaway that you get from this talk.
21:01And yes, please do connect with us.
21:03We'd love to talk if you have any questions.
21:06Thank you so much for listening.
21:08Yeah.