Self-hosted LLM agents reach frontier performance through harness optimization
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
A new method called CLEO enables government intelligence teams to deploy capable LLM agents on self-hosted, smaller models while maintaining security and auditability. Rather than optimizing the underlying model, CLEO tunes the agent harness, tools, memory, and control loop to match fixed infrastructure constraints. The approach achieves 83.8% accuracy on investigative analysis tasks, nearly matching Claude Code SDK performance while reducing system complexity by two orders of magnitude. This addresses a critical gap for security-conscious organizations that cannot rely on frontier model APIs.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
CLEO's core insight is that the agent harness (tools, memory, control flow, prompts) is the actual bottleneck for smaller models, not model capacity itself. The paper treats the underlying LLM as fixed infrastructure and reverse-engineers the deployment layer to compensate, which inverts the typical model-optimization playbook.
This directly extends the theoretical framework from 'Your Policy Is Secretly a Harness' (late October), which showed that harness structure encodes into learned policy and doesn't transfer across interface designs. CLEO operationalizes that insight for security-constrained environments: if the harness shapes what a model can express, then for teams locked into smaller self-hosted models, harness tuning becomes the only lever. The finding also echoes the Legora interview (this week), where privacy constraints forced novel engineering practices. Here the constraint is infrastructure rather than confidentiality, but the pattern is identical: work backward from what you cannot change and optimize everything else.
If government or defense contractors adopt CLEO for classified investigative workflows within the next six months and publish unclassified benchmarks showing the 83.8% figure holds on their own data, the approach has moved from research to operational validation. If adoption stalls or the accuracy gap widens on real-world classified tasks, the 2% loss versus Claude may be a dealbreaker for high-stakes intelligence work.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Agent policies encode their deployment harness as learned bias
A new theoretical framework reveals that language model agents absorb the structural biases of their deployment harness, the prompts and control logic that frame their tasks. Under KL-regularized optimization, researchers show that a trained policy encodes its harness's action preferences directly into its decision-making, meaning performance gains don't transfer cleanly across different interface designs. This…
MentionsCLEO · Claude Code SDK · Anthropic
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Optimize the Harness, Not the Model: Fitting a Minimal Agent to a Self-Hostable LLM for Investigative Analysis”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.