
01
Agentic systems
How coding agents plan, use tools and recover from mistakes: the work behind Line CLI and the Cosmos Agent API.
AI Research by Clatter Labs
We study models, inference and agentic systems, and publish what we learn, so builders can trust what they ship.
Research areas

01
How coding agents plan, use tools and recover from mistakes: the work behind Line CLI and the Cosmos Agent API.

02
Hallucination traps, unverifiable claims, sycophancy, pressure and injected instructions, tested in English and Egyptian Arabic.

03
Tokens, latency and cost per task measured as first-class results, not footnotes.
Approach
Every project starts with a real engineering problem.
Models are system components with inputs, tools, evaluation and operational boundaries.
We publish research, writing and much of our code.
The dotted line loops each iteration back into research.
Benchmark 3 · November 6, 2026
A first, small run. 2.0 is within one problem of 1.1 on coding and one to two cases behind on honesty and purity, while using more tokens and time per task. Small samples: treat it as a first comparison, not a ranking.
| Model | Honesty n=22 | Purity n=22 | HumanEval+ n=19 | MBPP+ n=15 | Cost / 1,000 tasks |
|---|---|---|---|---|---|
| SuperCosmos 2.0 Flash | 90.9% | 95.5% | 89.5% | 86.7% | $3.29 |
| SuperCosmos 2.0 Max | 86.4% | 95.5% | 94.7% | 86.7% | $5.55 |
| SuperCosmos 1.1 Flash | 100.0% | 100.0% | 94.7% | 86.7% | $1.76 |
| SuperCosmos 1.1 Max | 95.5% | 100.0% | 94.7% | 86.7% | $2.66 |
From research to products
01
BetaA very fast coding agent for your terminal, written in Rust. Slash commands, auto routing across the SuperCosmos models, skills, plugins and cloud delegation.
Explore Line CLI02
AlphaAn AI agent API for coding agents. Start a run, stream every step and get the result, with scoped keys.
Explore the Agent API03
Encrypted storage for code and assets, organized by project. Up to 10 GB per bucket.
Explore Code BucketsWriting & updates
Careers
Engineers, researchers, DevOps, designers and writers, fully remote. If nothing fits, write to us anyway.