Project AURIONAI Use-Cases Research & Innovation

Agents and Tool Use

Systems that plan, call tools and act over multi-step tasks.

Tier Grades

feasible

Tier A: Frontier AI labs

feasible

medium confidence

Frontier labs and NVIDIA (Nemotron 3 Nano, Super, Ultra for sub-agents) are observed building agent models.

feasible

Tier B: Indonesian AI labs

feasible

high confidence

A 100-line agent reached 65% on SWE-bench Verified, so agent systems need little training compute.

feasible

Tier C: Institution units such as ITB

feasible

medium confidence

Academic benchmark teams are observed and GAIA was built by 6 authors.

feasible

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

feasible

high confidence

A 100-line agent on existing models fits proof-of-concept scale, but the 7-month horizon doubling gives short shelf life.

Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.

Anchors

  • Computemini-SWE-agent reached 65% on SWE-bench Verified in 100 lines of Python; compute is mainly inference on existing models.
  • DataSWE-bench Verified 500 human-filtered instances; GAIA 466 questions.
  • Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): GAIA has 6 authors.
  • Time horizonFast moving; METR task horizon doubles about every 7 months, under 3 months on SWE-bench Verified.
  • Shelf-life signalsGAIA: GPT-4 with plugins 15% at release (Nov 2023), SWE-bench Verified: 65% with a 100-line agent by Jul 2025, METR horizon doubling about 7 months
  • Pipeline roleTier A ships the agent-capable models; Tier B and Tier C build agents and agent benchmarks on them; Tier D packages agents into proofs of concept and value, which age fast.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.

Who and What

Exemplars
WayveGAIASWE-benchNemotron agent modelsmini-SWE-agent
Observed actors
Frontier labsNVIDIANVIDIAAcademic benchmark teamsAnthropicAnthropic
Linked models
NVIDIANemotron 3LLMexemplarXiaomiMiMo-V2.6-Flashmultimodal LLMproduced bySakana AIFugu MaxLLMproduced bySakana AIFugu Ultra v2LLMproduced byAnthropicClaude Fable 5.1multimodal LLMproduced byinclusionAILing 3.0 Flash Fin (Ling-3.0-flash-Fin)LLMproduced byUpstageSolar Pro 4LLMproduced byMetaMuse Glimmer 30Bmultimodal LLMproduced byAlibaba (Qwen)Qwen3.8-27Bmultimodal LLMproduced byMotif TechnologiesMotif-3reasoningproduced byinclusionAILing 3.0 Flash (Ling-3.0-flash)LLMproduced byGoogle DeepMindGemini 3.6 FlashLLMproduced byMeituanLongCat-2.0LLMproduced byAlibaba (Qwen)Qwen-AgentWorld-35B-A3BLLMproduced bySakana AIFuguLLMproduced byMistral AIMistral Medium 3.5 (v26.04)multimodal LLMproduced byXiaomiMiMo-V2.5multimodal LLMproduced byAlibaba (Qwen)Qwen3.6-35B-A3BLLMproduced byArcee AITrinity-Large-Thinkingreasoningproduced byZ.ai (Zhipu)GLM-5V-Turbomultimodal LLMproduced byAlibaba (Qwen)Qwen 3.6 PlusLLMproduced byXiaomiMiMo-V2-Omnimultimodal LLMproduced byUpstageSolar Pro 3LLMproduced byKakaoKanana-2 30B A3B InstructLLMproduced by
+33
XiaomiMiMo-V2-FlashLLMproduced byAllen AIOlmo 3.1 32B InstructLLMproduced byDeepSeekDeepSeek-V3.2reasoningproduced bySpaceXAI (xAI)Grok 4.1 FastLLMproduced byOpenAIGPT-5.1LLMproduced byMicrosoft AIFara-7Bmultimodal LLMproduced byDeepSeekDeepSeek-V3.1-TerminusLLMproduced byAlibaba (Qwen)Qwen3-VL-235B-A22B-Instructmultimodal LLMproduced byMeituanLongCat-Flash (LongCat-Flash-Chat)LLMproduced byDeepSeekDeepSeek-V3.1LLMproduced byZ.ai (Zhipu)GLM-4.5Vmultimodal LLMproduced byAnthropicClaude Opus 4.1LLMproduced byAnthropicClaude Opus 4LLMproduced byOpenAIGPT-4.1LLMproduced byOpenAIGPT-4.1 miniLLMproduced byPerplexitySonar ProLLMproduced byHarvard UniversityTxAgent T1 Llama 3.1 8Breasoningproduced byMicrosoft AIMagma-8Bmultimodal LLMproduced byAnthropicClaude 3.5 Sonnet (Oct '24)multimodal LLMproduced byZ.ai (Zhipu)GLM-4-9B-ChatLLMproduced byCohereCommand R+ (Apr '24)LLMproduced byZ.ai (Zhipu)CogAgent chatmultimodal LLMproduced byZ.ai (Zhipu)CogVLM chatmultimodal LLMproduced byZ.ai (Zhipu)CogVLMmultimodal LLMproduced byAmazonNova ActLLMproduced byAnthropicClaude 3.7 SonnetLLMproduced byByteDanceSeed2.1LLMproduced byGoogle DeepMindantigravity-preview-09-2026LLMproduced byGoogle DeepMinddeep-research-max-preview-04-2026LLMproduced byGoogle DeepMinddeep-research-preview-04-2026LLMproduced byGoogle DeepMindgemini-2.5-computer-use-preview-10-2025LLMproduced byInternScienceAgents-A1-4BLLMproduced byLG AI ResearchEXAONELLMproduced by

Papers

Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.

Sources

  1. Measuring AI Ability to Complete Long Software Tasks metr.org https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
  2. GAIA: a benchmark for General AI Assistants arxiv.org https://arxiv.org/abs/2311.12983
  3. SWE-bench Leaderboards swebench.com https://www.swebench.com/
  4. NVIDIA Nemotron nvidia.com https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/
  5. Agentic misalignment: How LLMs could be insider threats anthropic.com https://www.anthropic.com/research/agentic-misalignment

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Agents and Tool Use", https://aurion.semai.id/topics/agents-and-tool-use/, records of October 2026.