# Coding Agents

Tencent has open-sourced LoopForge, a self-hosted workspace where a team describes a requirement to Main Bot in chat and it opens a Docker sandbox, checks out repositories, delegates to project roles and opens the PR. Previously, asking a coding agent to fix code left the result dependent on that run, which might skip the plan, forget tests or review its own code. LoopForge saves a proven way of working as an SOP, a standard workflow of agents, scripts and human gates, so later requirements follow the same stages and checks, with approvals, node status and tool traces kept in one place.

Tencent has open-sourced LoopForge, a self-hosted workspace where a team describes a requirement to Main Bot in chat and it opens a Docker sandbox, checks out repositories, delegates to project roles and opens the PR. Previously, asking a coding agent to fix code left the result dependent on that run, which might skip the plan, forget tests or review its own code. LoopForge saves a proven way of working as an SOP, a standard workflow of agents, scripts and human gates, so later requirements follow the same stages and checks, with approvals, node status and tool traces kept in one place.

BioXArena-Pro replaces biomedical modeling evaluation with 64 research-level tasks rebuilt from published studies, most holding out an entire biological unit such as a species, cohort or cell line, and audited in several rounds to close label leaks. Across 22 agent configurations the best scores 55.1 of 100 and no task saturates; on the nine tasks close to the original studies the strongest agent ties the reproduced published methods (54.9 versus 54.7), while on the 48 rebuilt tasks the methods trail by 8.2 points.

BioXArena-Pro replaces biomedical modeling evaluation with 64 research-level tasks rebuilt from published studies, most holding out an entire biological unit such as a species, cohort or cell line, and audited in several rounds to close label leaks. Across 22 agent configurations the best scores 55.1 of 100 and no task saturates; on the nine tasks close to the original studies the strongest agent ties the reproduced published methods (54.9 versus 54.7), while on the 48 rebuilt tasks the methods trail by 8.2 points.

AI agents can now propose atom-level biological hypotheses faster than they can be tested, but trust in an agent-run molecular dynamics simulation rests on a process record, and existing agents leave that record to the language model itself. MDClaw pairs validating tools with a directed acyclic graph (DAG, a graph recording which calculation stages depend on which) that checks stages before execution and records each one; on MDDataBench, which grades against deposited simulations, four open-weight models with MDClaw passed all three runs on 81–100% of 31 test systems. An ablation without the DAG found 11 runs that passed every output check yet had skipped minimization or equilibration, or simulated a different structure.

AI agents can now propose atom-level biological hypotheses faster than they can be tested, but trust in an agent-run molecular dynamics simulation rests on a process record, and existing agents leave that record to the language model itself. MDClaw pairs validating tools with a directed acyclic graph (DAG, a graph recording which calculation stages depend on which) that checks stages before execution and records each one; on MDDataBench, which grades against deposited simulations, four open-weight models with MDClaw passed all three runs on 81–100% of 31 test systems. An ablation without the DAG found 11 runs that passed every output check yet had skipped minimization or equilibration, or simulated a different structure.

Coding agents mostly append to their history, so prefix-cache reuse reaches 98–99%, yet text-level context compaction rewrites earlier tokens and forces the whole suffix to be recomputed. ObsDrop moves compaction from the prompt into the KV cache: it keeps the logical token prefix unchanged while releasing the KV of stale tool observations and excluding those positions from attention, implemented in SGLang and Sparse-vLLM. At 16 concurrent agents it reaches 2.45–2.85× throughput with FlashInfer and a 5.37× peak with Triton relative to no compaction, with task quality close to no compaction.

Coding agents mostly append to their history, so prefix-cache reuse reaches 98–99%, yet text-level context compaction rewrites earlier tokens and forces the whole suffix to be recomputed. ObsDrop moves compaction from the prompt into the KV cache: it keeps the logical token prefix unchanged while releasing the KV of stale tool observations and excluding those positions from attention, implemented in SGLang and Sparse-vLLM. At 16 concurrent agents it reaches 2.45–2.85× throughput with FlashInfer and a 5.37× peak with Triton relative to no compaction, with task quality close to no compaction.

Frontier models' spatial understanding now lets coding agents control a real robot arm zero-shot and end-to-end, with no task-specific demonstrations or retraining, so risk crosses from the digital world into the physical one. ArmedAgent sorts these risks by source into benign instructions that turn unsafe, malicious use, and indirect prompt injection, across 36 tasks in 5 scenarios, evaluated in simulation and validated on a physical arm. Testing GPT-6-Astra, GPT-6-Sol and GPT-6.1-Terra under Codex, overall attack success ranges from 72.22% to 94.29%, and refusal rate is 0% for every model in every category.

Frontier models' spatial understanding now lets coding agents control a real robot arm zero-shot and end-to-end, with no task-specific demonstrations or retraining, so risk crosses from the digital world into the physical one. ArmedAgent sorts these risks by source into benign instructions that turn unsafe, malicious use, and indirect prompt injection, across 36 tasks in 5 scenarios, evaluated in simulation and validated on a physical arm. Testing GPT-6-Astra, GPT-6-Sol and GPT-6.1-Terra under Codex, overall attack success ranges from 72.22% to 94.29%, and refusal rate is 0% for every model in every category.

Mellum2.1 is the next version of Mellum2, with the architecture unchanged: still a 12B mixture-of-experts model with 2.5B active parameters under Apache 2.0, with the changes concentrated after pre-training. Mellum2 was fast but could not work inside a repository at the level its developers wanted; this version, trained with reinforcement learning in real environments across thousands of environments and millions of sandboxed runs, can explore a codebase, edit files and check its own changes. Under one shared evaluation setup it is compared with Mellum2 and with the similar open models Qwen3.5-9B and Gemma 4 E4B, and the largest gain is in agentic coding, with improvements also reported in coding, competitive programming, math, tool calling and general knowledge.

Mellum2.1 is the next version of Mellum2, with the architecture unchanged: still a 12B mixture-of-experts model with 2.5B active parameters under Apache 2.0, with the changes concentrated after pre-training. Mellum2 was fast but could not work inside a repository at the level its developers wanted; this version, trained with reinforcement learning in real environments across thousands of environments and millions of sandboxed runs, can explore a codebase, edit files and check its own changes. Under one shared evaluation setup it is compared with Mellum2 and with the similar open models Qwen3.5-9B and Gemma 4 E4B, and the largest gain is in agentic coding, with improvements also reported in coding, competitive programming, math, tool calling and general knowledge.