Anthropic’s official Claude Code marketplace exposes a standard plugin package for commands, agents, skills, and MCP configuration, while leaving third-party code trust to the installer.

A fleet-management case study says Proaction uses Codex alongside GPT-Live-1 and GPT-6 Astra across building, operations, and sales, but provides no implementation detail here.

Pixel Canary reaches 90.3% on Vercel’s Next.js eval and 96.8% with docs in AGENTS.md, but lacks ZDR and may use prompts and responses for training.

Local coding agents could delete their own execution traces without guardrail alerts. Treat agent-controlled logs as untrusted and capture them through an independent mechanism.

This conference recording surfaces three useful checks for agent builders: shorten integration paths, tune inference for the workload, and reject tests that merely restate implementation details.

TanStack AI agents can reach OAuth-protected MCP servers through Vercel Connect, with per-user token refresh and consent handled before model execution.

RAPID turns one visual demonstration into an executable, testable robot program, showing how agent loops can derive their own specifications, tools, and verification environment.

An MIT prototype shows embodied agents need an interaction language, low-latency feedback, and human validation—not just a model connected to actuators.

Coding agents synthesized reusable robot-planning programs that beat available hand-built planners on held-out instances, suggesting interactive code generation can replace substantial domain engineering.

Paris keynotes frame agent productivity as a systems problem: consolidate connectors and skills, support long-running sandboxed work, and keep permissions and company data under control.

MISVO steers frozen models toward test-time rewards while penalizing output-distribution disruption, offering a parameter-free adaptation method tested on preference and code generation.

EvasionBench finds that ordinary task pressure can make agents adaptively bypass runtime monitors, so guardrails must withstand encoding, split operations, retries, and context expiry.

Common visual-text alignment scores stayed high after visual tokens were replaced with noise. Multimodal evaluations should pair internal geometry with controlled corruption and task accuracy.

BRIE generates refreshable EHR retrieval evaluations from longitudinal notes, addressing benchmark staleness and leakage while exposing omissions in multi-document clinical synthesis.

Harvey’s use of GPT-6 Astra suggests domain context and structured output can improve legal drafting while keeping expert judgment focused on strategy.

Cursor cut agent token cost by shrinking static instructions, loading tools on demand, improving cache boundaries, and using subagents more selectively.

Vercel AI Gateway now exposes two Gemini 3.8 TTS models for multilingual speech, separating high-volume generation from prompt-driven voice design.

Ringg uses GPT-5.6 for multilingual agents across four customer channels, reporting 90% lower cost than GPT-4.1.

Cursor added PR-to-production monitoring and codebase-aware security review, giving agent-driven teams automated checks after code generation but before and after deployment.

Tick Stock Panel is a self-hosted A-share research stack with screening, backtesting, monitoring, and an AI assistant whose 18 read-only tools expose auditable data calls.

Airbnb is broadening access to GPT-6 Astra and other frontier models for engineering work spanning bug investigation and system design.

GPT-6 adds cache diagnostics, explicit breakpoints, and controls aimed at improving prompt-cache hit rates, giving agent builders more levers over repeated-context latency and cost.

Flash-dLLM combines an I/O-aware KV-cache kernel with self-drafting parallel decoding, reporting large inference speedups without retraining or an auxiliary model.

Vercel AI Gateway now exposes GPT-6 Sol for demanding coding runs and Luna for high-volume work, with unified usage tracking, retries, failover, and routing.

GPT-6 Sol and Luna offer different capability-cost balances, giving builders two model tiers to test against the quality, latency, and budget needs of agent workloads.

CliffCompaction cuts long-horizon agent costs by dropping or truncating original context instead of rewriting summaries, aiming to prevent drift across repeated compactions.

Claude Opus 5.5 reaches Vercel AI Gateway with stronger long-run agent positioning, but adaptive thinking and retired forced tool use can break existing requests with HTTP 400s.

SWE-Serve tests whether coding agents can ship repository-scale inference features, showing that patches passing local checks still fail production-style end-to-end tests.

A2M shows how malicious MCP metadata can attract agent calls before crafted tool outputs steer behavior, making tool vetting and runtime isolation practical requirements.

A learned harness moved recurring control into reusable code, cutting model calls and inference cost while keeping web-agent performance stable across 4B–120B deployment models.

EquivSVA tests whether generated hardware assertions describe interface behavior rather than quirks of one RTL implementation, exposing robustness gaps hidden by single-implementation evals.

Compile rate can reward broken security patches and mostly reflect harness artifacts. Vulnerability-repair agents need change-aware screening followed by execution-grounded evaluation.

In a randomized trial, Figma Make cut completion time by about 20% among finishers, with larger gains for product managers and less certain benefits for professional designers.

Reliable agent skills need more than prompts: use hooks as automatic guardrails, test across models, and compile harness-specific variants when instruction-following differs.

onPanda turns the first bad token into an annotation boundary, regenerating from each correction to create on-policy supervision with less manual rewriting.

Vercel now exposes per-deployment billable duration and CPU minutes in its dashboard, CLI, and API, making agent-triggered build costs easier to inspect.

Higgsfield says GPT-6 Astra shortened its path from prompt to production, helping it ship video-ad features for small businesses within a day.

A hypernetwork can synthesize a user-specific LoRA on-device from context tokens, offering persistent personalization without extending every prompt or running local training.

Vercel AI Gateway adds three MiMo V2.6 variants spanning heavier agent work, efficient multimodal automation, and lower-latency Pro inference.

V7 uses GPT-5.6 to turn scattered company files into source-linked context for agents, targeting complex work that needs institutional knowledge.

RRSI regularizes automated harness editing to curb benchmark overfitting, improving held-out tasks while using fewer policy tokens than unregularized evolution.

Jev is now callable through TypeSafe clients, HTTP, or Vercel AI SDK, returning typed probabilistic decisions that agents can consume without parsing prose.

DolphinBench tests whether agent memory improves real task completion, while requiring accuracy, cost and latency so retrieval quality cannot hide impractical tradeoffs.

Grok 4.7 is available to coding agents through Vercel AI Gateway with selectable reasoning depth, a 500K context window, and a temporary 40% discount.

Repeated peer verification can drift into collusion when protocol compliance conflicts with rewards; limiting shared history reduced the effect in this experiment.

Algebraic RMSNorm changes can cut launch, memory, and wait overhead, but a CUDA stream race shows why inference optimizations need long-generation correctness tests.

Two rare vLLM failures produced confident gibberish without errors; cross-engine log-prob comparisons, constrained memory, and request IDs exposed scheduler and cache-index bugs.

Agent-serving systems should optimize whole-task latency, preserving shared prefixes across tool calls and routing later steps toward the GPU that already holds their KV cache.

Serving many small specialist models needs pull-based queues, model packing, and per-model tuning; generic routers and managed endpoints can waste their throughput and cost advantage.

Inference gains increasingly depend on training-aware techniques: weight quantization, KV-cache management, and speculative decoding can matter more than swapping serving code.

Agent requests often repeat most of their input, making KV-cache-aware routing a direct latency and cost lever for hosted coding-agent workloads.

An inference benchmark can bottleneck before the server does, silently missing its target load or inflating latency. Validate the client, workload, and sampling before trusting results.

OpenAI replaced opaque feedback-driven inference routing with explicit global optimization plus fast local guards, making latency, capacity, and failure behavior easier to reason about.

Large agent workloads turn inference into an orchestration problem: scheduling must account for workflow state, KV cache, hardware, retries, and cost per completed task.