The Unbearable Lightness of Agent Optimization — Alberto Romero, Jointly

Right. Hello everyone. Uh, today I will present Meta-Adaptive Context Engineering, or Meta-ACE for short, which is a new framework designed to optimize AI agents beyond single-dimensional approaches. We will explore how orchestrating multiple adaptation strategies can overcome the limitations of existing context engineering methods.

Now, a little introduction about myself. So, I'm Alberto Romero, I'm the Co-founder and CEO at Jointly. And for context, at Jointly we build domain-specialized agents for regulated industries where policy adherence constraints are particularly strict. Most of our research work is in the area of self-optimizing agent architectures, uh, using systematic approaches.

Now, about myself: I have spent 20-plus years at the intersection of AI and data. Some of my recent experience includes being the CTO and Co-founder of Human.ai, a think ML-based risk prediction for mobility, which was acquired by Aon in 2023.

And in my previous role, I headed up Citibank's Gen AI engineering team. Now, here's our agenda for today. We'll begin with the motivation and problems that current systems face, then we'll review the ACE framework and its limitations. After surveying recent research insights, we'll introduce the Meta-ACE approach, we'll discuss its architecture and strategy toolbox, show some results, and finish with future directions and challenges.

Now, the Agentic Context Engineering Framework, or ACE for short, for which you've got the paper link on the slide there, is a very popular framework, and the paper came out a few months ago. It basically organizes adaptation into three roles.

First of all, there's a generator that produces reasoning paths, then there's a reflector that extracts lessons, and finally there is a curator that synthesizes these lessons into incremental updates. ACE uses incremental delta updates and a grow-and-refine mechanism to prevent context collapse and maintain relevance.

Now, most importantly, it can improve without labeled data by learning directly from execution feedback.

Now, so, ACE has been quite successful and has achieved substantial gains across some of the most popular agent benchmarks, like UpWorld or Finer, with almost an 11% compared to previous state-of-the-art approaches such as JEPA or DC.

And it's also achieved an 8.6% gain on financial reasoning tasks.

There are four fundamental limitations for ACE that I'm going to reflect on and just discuss on the next slide, and those form the basis for Meta-ACE, basically. Now, as I was saying, despite its strengths, ACE has got four critical failure modes.

First, it is highly dependent on the reflector, so when reflection fails, the context becomes noisy and even harmful. Secondly, there's feedback brittleness, which means that when ground truth signals are weak or absent, ACE may reinforce incorrect behaviors. Third, there's the task complexity blindness, which leads to treat simple and complex tasks the same, which can be a waste of resources and also a miss of opportunities for optimization.

And then finally, ACE optimizes only the context dimension, so it ignores compute, memory, and parameter updates.

Now, the 2024 and 2025 research landscape offers four key insights, in my views. First of all, verification mechanisms, like self-evaluation, multimodal consensus, and execution checks, are really important for robustness of any solution. Secondly, adaptive compute allocation shows that small models can outperform much larger ones by selectively increasing inference steps.

The third one is that structured memory architectures outperform linear context accumulation by organizing facts as graphs or multicranial memories. And then finally, test-time training bridges inference and learning, and enables temporary parameter updates to yield large accuracy gains. So, these advances suggest that we need a hybrid multidimensional system.

Now, Meta-ACE addresses ACE's limitation by adding a meta-controller that learns to orchestrate multiple adaptation strategies based on a task's complexity, uncertainty, verifiability, and also resource constraints. So instead of applying the same procedure to every problem, Meta-ACE profiles each task and allocates theright combination of strategies across context, compute, verification, memory, and parameter dimensions.

So this adaptive learned coordination is what enables it to outperform single-dimensional methods.

Now, the Meta-ACE framework consists of four layers, so getting into the architecture. The first layer is the task profiling one, which assesses complexity, uncertainty, verifiability, and resource budgets. Then there is a lightweight meta-controller that selects and allocates adaptation strategies accordingly.

The next layer down is the strategy execution one, and it carries out the reflection, adaptive compute, hierarchical verification, structured memory retrieval, and selective test-time training. And then finally, there's a feedback aggregation layer that collects the outcomes and updates the meta-controller's policy through meta-learning.

So this layer design allows the system to learn from its experience and continuously refine its decision-making.

Now, in terms of the task profiling, there are four key dimensions that are being assessed. The first one is semantic complexity. So this is basically an embedding-based similarity to known task distributions that gets produced. The second one is uncertainty quantification.

Think of it as a relative softmax scoring that predicts model confidence. The third one is verifiability assessment, so whether we can execute and validate the output. And then the fourth one is resource availability. So we take into consideration the context window, the compute budget, and even other constraints such as time.

So the output of this layer, of the task profiling layer, is a 32-dimensional task embedding, which is what feeds as input into the meta-controller.

Now, in terms of the strategy toolbox, Meta-ACE draws from six strategies. First one is minimal context, which uses concise prompts for simple tasks. Then we use ACE reflection, which retains the generator-reflector-curator loop for incremental knowledge accumulation, as established by standard ACE.

Then we also use adaptive compute, which scales the number of reasoning steps or samples based on the task difficulty. We also use hierarchical verification that combines self-evaluation, multimodal consensus, and execution checks.

Adaptive memory that retrieves relevant information from structured multicranial memories. And then finally, we use selective test-time training, which applies temporary parameter updates, such as LoRA adapters, for high-stakes tasks. So the meta-controller learns to combine these tools effectively over time.

Now, the reward formula upon which the learning strategy is selected accounts for the following components. The first one is the correctness of an action or prediction, which is accuracy. Then we also have the penalty associated with resources used or negative outcomes, so 1 minus cost.

And then is the trustworthiness of the models, which is self-expressed certainty, so the confidence calibration, basically, with weighted importance determined by the hyperparameters alpha, beta, and gamma.

In terms of the meta-learning loop, we have four sources of feedback collection. First of all is task outcomes, the success, failure, or correctness of the task. Then we've got the strategy performance, so what is the individual contribution of each strategy to the overall performance of the task.

Then we also have efficiency metrics, such as the compute, latency, memory. And then finally we've got confidence calibration, so where predictions are accurate.

So, moving on to how we go on about solving the limitations from ACE. The first one was the weak reflector problem. So ACE's issue is that there is a 50 to 60% performance drop when reflector quality degrades. With Meta-ACE, we introduce three things, basically.

So first of all is quality gates, so it's a learned classifier that blocks harmful deltas. And secondly, there's a multi-signal reflector, or reflection, which basically is an ensemble of specialist models when there is a level of uncertainty. And then the third one is adaptive strategy allocation.

So the meta-controller learns when reflection fails, and then it routes to verification or test-time compute instead. So we can expect to maintain an 80%-plus performance even when the reflector degrades around 30%.

Now, the second limitation we had was the feedback quality brittleness. So what we observe with ACE is that there can be significant degradation without reliable ground truth signals. With Meta-ACE, we introduce a hierarchical verification cascade, where we can expect a 50 to 60% reduction in errors from poor feedback, and that's through three tiers.

The first tier is self-verification, which is just a fast filter, which just accepts if the confidence level is over a certain value. Second tier is multimodal consensus, so we leverage a diverse range of models, such as GPT-4, Claude, and DeepSeek, and we do confidence-weighted voting.

And then the tier three is execution-based verification, where we leverage code sandbox, API validation, and schema compliance.

The third limitation we had was task complexity mismatch. So in a sense, the fact that ACE uses uniform processing also for simple tasks, which can be a waste of resources. So Meta-ACE adapts strategy allocation dynamically rather than using the same heavy pipeline for everything.

The alphas are allocation weights for the six optimization strategies, and they represent how much computational budget is assigned to each strategy for a given task. So simple tasks require minimal processing, can save around 90% compute compared to standard ACE.

Moderate tasks is more of a balanced approach that includes ACE plus verification. And then complex tasks, basically heavy test-time compute, multiple attempts, and memory retrieval.

So just to conclude with some results, and these are initial results, we have observed around an 8 to 11% improvement on agent benchmarks. We have also observed a 6 to 8 points improvement on some domain-specific tasks. Also a 30 to 40% reduction in compute costs through the allocation of adaptive strategies.

And overall, there's more robustness, more consistency, and we can generalize better. We can use the framework across a diverse range of domains. So the conclusion is that Meta-ACE can orchestrate context compute and verification and memory and parameter adaptation, and produce a robust self-improving framework for agents.

Future work will implement and evaluate the full system across

a more diverse range of domains, and we'll continue exploring meta-learning, and this will involve also incorporating additional strategies as well.

Now, I also wanted to touch on additional applications of Meta-ACE that I think are quite relevant. So first one is for multimodal AI systems. So, for example, deciding when to use vision versus language processing, again, can be

a meta-adaptive strategy decisioning. Also, when you have compound AI systems that require different models for different stages, and the complexity is

substantial, we can actually, in a meta-adaptive manner, select the most effective strategies to resolve a task end-to-end. Also, for human collaboration, so in other words, to determine when to have a human in the loop. And also for continual learning systems, where we are balancing exploration versus exploitation.

So the core takeaway is that optimization requires a meta-layer of intelligence, and that has to be trained. And

it requires a lot of trial and error before it can actually perform at theright level. In terms of the future direction and challenges, there are still several challenges that remain. So the meta-controller's training may be unstable due to sparse rewards, and this can be mitigated through curriculum learning.

Also, robust advantage estimation and regularization of entropy. Also, computational overhead from profiling and multiple strategies needs to be reduced with efficient models. We can leverage things like lazy execution, batching, and caching.

Also, the verification cascades can be brittle if all models make the same mistake. So we need diverse models with confidence weighting and human oversight, as well as active learning. Meta-learning loops require substantial data, synthetic task generation, or policy learning transfer from related domains, and simple efficient algorithms can also help as well.

And finally, addressing these challenges is going to be key to scaling Meta-ACE and applying it across a wide range of domains. So that was all from me. Thank you very much for listening. And yeah, I appreciate you being there.

Thank you.