Today’s lead · X
SemiAnalysis opens AgentX for long-context agentic inference benchmarking
SemiAnalysis says AgentX 1.0 brings a fully open-source, multi-turn agentic coding inference benchmark at 1M context under Apache 2.0. The useful shift is measuring messy production patterns: prefill reuse, sub-agent bursts, KV cache offload, tool calls, and hardware/software stacks beyond fixed sequence benchmarks.
Tools & repos
4 selectedapache/maka
Apache Maka is a local-first AI agent workspace that records model messages, tool calls, tool results, permission decisions, and termination events as an append-only log. That audit trail is the real builder hook.
FetchSandbox MCP
FetchSandbox MCP targets a painful gap in agentic coding: an integration fix that passes CI but still returns wrong data. It claims 70+ API sandboxes and one config block for Cursor or Claude Code.
Construct Computer
Construct Computer pitches an AI workforce for solo founders and small teams. The promise is practical if it holds: install MCPs or skills like apps, let agents build missing tools, and turn useful actions into reusable workflows.
freestylefly/awesome-gpt-image-2
awesome-gpt-image-2 is a prompt-as-code repository for GPT-Image2, with 470+ reverse-engineered cases and 20+ industrial templates. Treat it as a working prompt library, not a benchmark.
Blogs worth your time
0 readsNo long-form read was strong enough to recommend today.
Community discussions
4 threadsCloud agents are turning local dev into the bottleneck
The thread argues worktrees fail under real stacks because databases, ports, and dev servers collide. The proposed pattern is one cloud computer per agent, with live portals for review and heavier proof runs than a laptop tolerates.
Dual RTX 3090s can serve many local agents, but TTFT bites
This local inference sweep is useful because it shows where consumer GPUs actually bend. On 2× RTX 3090s, aggregate throughput peaked around 306–308 tok/s, but TTFT rose from 0.6s to about 12s at 32 streams.
RAM boots giant models; VRAM decides usable context
The thread corrects a common local-model buying shortcut: system RAM may load a huge checkpoint, but VRAM and KV cache decide whether the context window is practical. The reported GLM-5.2 run loaded, yet only decoded at 7.5 tok/s.
Cursor users debate quality drops and enterprise-grade support
A frustrated Cursor user reports crashes, failed refactors, and deleted code after an update. The comments reveal a sharper split: enterprise admins describe responsive escalations, while individual users complain about poor support and Discord moderation.
Funding & acquisitions
0 movesNothing material cleared the editorial bar today.
Bengaluru radar
2 events
The Agent Autopsy - Breaking and Fixing AI Agents
A free in-person Bengaluru workshop on agent security failures, with a compromised demo agent plus red-team/blue-team sandbox exercises.
Hands-On: Build Agentic Workflows and Searchable Apps with Elasticsearch, Jina, and Agent-to-Agent Comm.
A free Bengaluru code-along workshop for building agentic search apps with Elasticsearch, Jina embeddings, Elastic Agent Builder, and A2A.