Fork sandboxes from snapshots with inherited config/env via `Sandbox.fork()`, overriding specific fields as needed.
Enables parallel sandbox variants and multi-tenant isolation without manual config duplication. Cuts setup time for branched agent workflows and template distribution.
Replaces manual sandbox cloning. Requires SDK upgrade to @vercel/sandbox@latest or CLI to sandbox@latest. Ready now—SDK and CLI both ship the feature today.
“The fork starts from the source's current snapshot and inherits its config and environment variables”
“A fork takes about the same time as creating a sandbox, with the same limits”
“If the source has no snapshot, it falls back to a fresh create, using the source's runtime and config”
“Any parameter you pass overrides the inherited value”
vercel-sandboxisolationsdkcliagent-dev
Dev Signal
Get issues like this in your inbox — free, every weekday.
Quick Signals
Kimi K3 opens 2.8T agentic model with million-token context
Open-weight 2.8T MoE model with native vision, 1M token window, and 16-of-896 expert routing achieves competitive coding and agentic task performance at ~2.5× scaling efficiency over prior version.
Developers gain access to frontier-class weights for local deployment without licensing friction. Long-context window and native multimodal support enable complex code navigation, repo-scale refactoring, and vision-in-the-loop workflows without external API calls.
Replaces closed API dependency for long-context coding tasks if you can host 2.8T parameters. Requires VRAM for 104B activated params (MoE sparsity), quantization support (MXFP4 weights). Worth evaluating now if you have GPU infrastructure; benchmark gaps vs. Claude Fable 5 on reasoning (CritPt: 23.4 vs 28.6) and some coding tasks suggest it's not a universal replacement yet.
“2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window”
“activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2”
“release the full Kimi K3 model weights under the Kimi K3 License”
Set dollar limits per team, project, or API key; requests rejected when any budget hits its cap.
Prevents runaway AI API costs at organizational scope rather than just per-key granularity. Multi-level budget enforcement means a single runaway project can't drain team allocations.
Data Point
DNS-based tool discovery cuts agent search O(N) to O(log N)
ToolDNS embeds semantic intent into DNS hierarchies, replacing HTTP registries with O(log N) lookups and 95% smaller search spaces via UDP.
Autonomous agents querying millions of tools hit wall-clock latency and centralized governance bottlenecks. DNS-native discovery eliminates middleware overhead, shifting tool routing from semantic search to deterministic name resolution—critical for production agent deployments.
Replaces HTTP-based tool registries and semantic search indexes for agents. Requires DNS extensions (EDNS0, partially unfolded names) and protocol adapters (MCP, RESTful, Skill). Research artifact with benchmark (33,688 tools) but no public implementation—evaluate for agent infrastructure pilots only, not production yet.
“ToolDNS slashes the per-query search space by 95.26% while matching state-of-the-art retrieval accuracy”
“O(log N) name resolutions”
“UDP-native design reduces discovery latency by orders of magnitude compared to HTTP-based registries”
“33,688 real-world tools spanning MCP, A2A, RESTful, and Skill protocols”
Replaces manual spend tracking across distributed teams. Requires setting budgets via dashboard or CLI with refresh periods (daily/weekly/monthly/cumulative). Ready now—native to Vercel AI Gateway with email alerts at 50%, 75%, 100% thresholds.
“Set a dollar limit on a scope, and the gateway meters spend against it and stops further requests once the limit is reached, until the budget resets or you raise it”
“A request can fall under multiple budgets and has to pass all of them: if any one is over its limit, the request is rejected, even when the others have room left”
“BYOK spend is not counted against budgets by default”
cost-controlai-gatewaybudgetsvercelobservability
AI Gateway logs page lists every request routed
Dedicated logs UI shows cost, token counts, routing attempts, and fallback paths for each gateway request—drill into failed attempts to see provider timeouts or budget exhaustion.
Observability into multi-provider routing decisions and cost attribution per request eliminates guesswork when debugging why a fallback triggered or which provider handled a request. Token breakdowns across input/output/cache let you optimize expensive modalities.
Replaces manual log parsing or CloudWatch-style dashboards for AI Gateway traffic. Requires no setup—available now at team and project scope if you use Vercel's AI Gateway. Worth checking immediately if you route across multiple providers or need per-request cost visibility.
“listing every request your team sends through the gateway, newest first, with cost, token counts, duration, and the model, provider, and region that served it”
“Click any row to open its details next to the list: total cost, a token breakdown across input, output, reasoning, and cache reads and writes, duration, time to first token”
“The fallback path lists every provider attempt for the request in order. Each attempt shows the model, provider, status code, credentials, and how long it took, and failed attempts include the reason, such as a provider timeout or an exhausted routing budget”
Poolside's Laguna S 2.1 now handles 10x request volume on Vercel's AI Gateway, enabling higher-throughput agentic coding without provider switching.
Developers running high-volume coding agents or long-running tasks hit capacity limits faster. This removes that friction for both free and paid tiers without API changes.
Drop-in model swap: set `model` to `poolside/laguna-s-2.1` or `poolside/laguna-s-2.1-free` in AI SDK calls. No infrastructure changes required. Worth testing now if you're agent-heavy or hitting rate limits.
“10x more capacity for Laguna S 2.1 on AI Gateway”
“good for high-volume agentic coding and long-running tasks”
“AI Gateway gives you one API to hundreds of models, with usage tracking, retries, failover, and higher-than-provider uptime built in”
“It reflects provider pricing with no markup and no platform fee”
laguna-s-2.1ai-gatewaycoding-agentscapacityvercel
Agent Runs surface subagent activity on eve projects
Subagents tab in Agent Runs shows delegated agent execution with prompt, duration, failures, and shared timeline—drill into any subagent run for full observability.
Debugging multi-agent workflows now requires fewer context switches: you can trace delegation chains, spot subagent failures, and correlate token usage across hierarchies without jumping between dashboards.
Replaces manual log inspection for subagent behavior in eve projects. Requires no code changes—available now in Agent Runs dashboard. Worth enabling immediately if you use delegated agents.
“You can now inspect subagent activity for eve projects in Agent Runs”
“The new Subagents tab shows every subagent, organized by which turn started it”
“Each row shows the prompt, duration, and any failures, all on a shared timeline”