SIA: The Self-Improving AI Agent That Updates Its Own Weights

By Prahlad Menon 3 min read

Most AI agents are frozen the moment they ship. Their scaffolding, weights, and memory never change again. A new paper from Hexo Labs breaks that assumption entirely.

The Bottleneck Is Us

Every AI system today—whether a fine-tuned LLM or an agentic wrapper—is improved by humans. We write the prompts. We tune the weights. We debug the retry logic. The long-horizon goal of an AI that can figure out how to improve itself has remained stubbornly open.

Two research lines have attacked this bottleneck in isolation:

  1. Harness-update school: A meta-agent rewrites the scaffold (tools, prompts, search procedures) while model weights stay fixed
  2. Test-time training school: RL pipelines update model weights on task feedback while the harness stays fixed

These silos operate in isolation. SIA (Hebbar et al., 2026) proposes doing both at once.

How SIA Works

SIA coordinates three agents in a self-improvement loop:

  • Meta-Agent: Reads the task description and generates an initial Target Agent
  • Target Agent: Attempts the task and logs its actions
  • Feedback Agent: Reviews performance logs, identifies improvements, and updates both the harness and the weights

The feedback agent doesn’t just tweak prompts—it triggers LoRA fine-tuning on task failures, building domain intuition that no prompt engineering can instill.

The Results

On three wildly different benchmarks:

TaskBaselineSIA-W+HImprovement
LawBench (191 Chinese legal charges)45%70.1%+56.6%
AlphaFold-3 TriMul GPU kernel1,161 μs1,017 μs14× faster
scRNA-seq denoising0.220 MSE0.289 MSE+31.4%

But here’s the wild part: on OpenAI’s MLE-Bench Hard—a gauntlet of real Kaggle ML competitions—SIA reached #1, then dethroned its own earlier versions.

Why This Matters

The key insight: harness updates and weight updates are complementary, not competing.

  • Harness updates make the model agentic—shaping how it searches and acts
  • Weight updates build domain intuition that no scaffold can instill

When SIA fails on Chinese legal cases, the feedback agent doesn’t just add “think step by step” to the prompt. It fine-tunes the model on the failure cases, building implicit knowledge of legal precedent that emergent in-context learning can’t match.

Try It Yourself

SIA is MIT-licensed and ships with four bundled tasks:

pip install 'sia-agent[claude]'
export ANTHROPIC_API_KEY="..."
sia run --task gpqa --max_gen 5 --run_id 1

The built-in web dashboard shows per-generation code diffs, improvement rationales, and accuracy-over-time charts.

The Question

If agents can now update their own weights, prompts, and tools in a closed loop…

What happens when we stop being the bottleneck?

How This Connects

SIA validates the direction we’ve been exploring in our own work:

autoloop — our open-source generalization of Karpathy’s autoresearch — runs a similar experiment loop on any file you want to improve. The key difference: autoloop currently only updates the harness (prompts, scaffolding), while SIA adds the weight update lever. The combination is where the magic happens.

Anthropic’s June 2026 disclosure revealed that Claude writes 80%+ of Anthropic’s production code. SIA takes this further — not just AI writing code, but AI deciding what code to write next based on its own performance feedback.

The gap is closing. Self-improving loops like SIA, autoloop, and Anthropic’s internal tooling are converging on the same pattern: agents that can update their own scaffolding, weights, and memory without human intervention.

Links:

Related posts: