← back
The Unbearable Lightness of Agent Optimization — Alberto Romero, Jointly
Takeaway
Single-axis context engineering hits a ceiling; multi-dimensional meta-controllers that pick the right tool per task profile beat uniform pipelines on both quality and cost.
Summary
- Jointly's Meta-ACE extends the ACE (Agentic Context Engineering: generator/reflector/curator) framework with a meta-controller that learns to orchestrate adaptation strategies across context, compute, verification, memory, and parameter dimensions.
- Identifies ACE's four failure modes: reflector dependency (50-60% perf drop), feedback brittleness without ground truth, task-complexity blindness, and single-dimension (context only) optimization.
- Task profiler outputs a 32-dim embedding (semantic complexity, uncertainty, verifiability, resource budget); strategy toolbox has 6 options including minimal context, ACE reflection, adaptive compute, hierarchical verification, adaptive memory, selective test-time training (LoRA).
- Hierarchical verification cascade (self → multi-model consensus weighted by confidence → execution sandbox) reduces poor-feedback errors by 50-60%; simple-task minimal-context saves ~90% compute.
agent-optimizationcontext-engineeringverification
Original description
This talk introduces Meta-ACE, a learned meta-optimization framework that dynamically orchestrates multiple strategies (context evolution, adaptive compute, hierarchical verification, structured memory, and selective test-time parameter adaptation) to maximize task performance under real-world constraints. Rather than relying on uniform prompt refinement, Meta-ACE profiles each task (complexity, verifiability, feedback quality) and selects an optimal strategy bundle via a lightweight meta-controller. Alberto is a seasoned AI and ML leader with over 20 years of experience at the intersection of AI and data. A hands-on engineer, Alberto has designed and built low-latency, mission-critical ML systems, and has specialized in systematic optimization of AI pipelines and agents using custom built evaluation techniques. He is an exited co-founder having sold his previous startup, Humn.ai, to Aon in 2023, which delivered real-time, ML-powered risk prediction for mobility. Alberto is the Co-founder and CTO at Jointly, specializing in self-optimizing AI agents for regulated industries. He holds an MSc in AI and Machine Learning and speaks at global AI conferences, including ODSC and AIAI. --- Socials: - LinkedIn: https://www.linkedin.com/in/albertoromero-uk/ - GitHub: https://github.com/a-romero - Company: Jointly (https://getjointly.ai)