AI agents, vector search, and isolated sandboxes — Dev Signal
AI agents, vector search, and isolated sandboxes
Share:
Tool of the Week
Hermes Agent integrates Vercel AI Gateway and Sandbox
Route Hermes inference through Vercel AI Gateway (200+ models, no markup) and execute agent commands in isolated microVMs instead of locally.
Centralizes model routing and spend tracking in one dashboard while isolating untrusted agent execution. Eliminates local filesystem risk when running autonomous coding tasks.
Replaces local model selection and direct shell execution. Requires Hermes install, Vercel account, and `hermes setup` reconfiguration; opt-in for Sandbox backend. Ready now—straightforward setup with `hermes doctor` verification.
“Access 200+ models through AI Gateway at no markup on tokens”
“Commands run locally until you set `terminal.backend` to `vercel_sandbox`”
“The backend supports `node24` (default), `node22`, and `python3.13` runtimes”
Get issues like this in your inbox — free, every weekday.
Quick Signals
DynamoDB enables native vector search without separate stores
Store embeddings directly in DynamoDB and query them with single-digit millisecond latency—eliminates ETL to specialized vector DBs.
Reduces operational complexity by collocating vector and relational data, cutting latency and infrastructure overhead for RAG pipelines and similarity search workflows built on existing DynamoDB deployments.
Replaces managed vector database overhead for AWS-native stacks. Requires schema refactor to add vector columns (up to 4,096 dimensions) and migration of existing embedding pipelines. Worth adopting now if you're already DynamoDB-heavy; monitor 100-result query limit for ranking workloads.
“supports up to 4,096 dimensions, Euclidean, Cosine, and Dot product distance functions, and returns up to 100 results per query with single-digit millisecond latency at 99% or higher recall”
“It is generally available in all commercial AWS regions, including the AWS GovCloud (US) regions”
dynamodbvector-searchembeddingawsrag
Cloudflare ships Kitesurf, browser optimized for AI agents
Kitesurf runs on Workers as a Wasm-compiled Rust engine, trading pixel-perfect rendering and human UI chrome for token efficiency and scalability—agents need cheaper browsers than Chromium.
AI automation workflows currently require full Chromium instances per agent, consuming prohibitive memory/CPU. Kitesurf on Workers enables cost-effective agent scaling by optimizing for machine-readable output over visual fidelity, directly reducing execution costs for agentic tasks at scale.
Data Point
MLLMs fail creative cross-concept reasoning tasks
C4 benchmark exposes 50.7% max accuracy on Chinese idiom decoding—current models struggle mapping non-obvious conceptual relations needed for generative tasks.
If you're building creative generation systems (design, education, human-AI collaboration), this reveals a concrete weakness: MLLMs don't reliably decode creative intent across abstract concept bridges. Your creative outputs will miss nuance your users expect.
This doesn't replace anything yet—it's a diagnostic. The C4-Eval framework itself (184 synthetic + 37 human items across 5 task settings) is useful for stress-testing your MLLM choice before committing to creative workflows. Worth running internal evals before shipping; results suggest no current model is production-ready for this class of task without heavy prompt engineering or fine-tuning.
“the strongest closed models reach 50.7% and 48.0% primary accuracy, while open-source models remain substantially lower”
“a substantial gap in how current MLLMs decode creatively encoded meaning through cross-concept relations”
“bridge hints and explanation requests provide only modest gains”
Replaces Chromium instances in Browser Run automation pipelines; requires Workers environment and acceptance of non-pixel-perfect rendering. Beta-free tier available now. Worth evaluating if your agents run screenshots/HTML extraction at scale—measure token savings against rendering accuracy tradeoffs first.
“browser engines like Chromium were built for humans, not agents, and they come with overhead that AI models simply do not need”
“Kitesurf is significantly more efficient in CPU and memory consumption than Chromium for common agentic tasks like screenshots and HTML extraction”
“AI doesn't care about tabs, themes, browser extensions, or synchronization across devices. It cares about token count, context windows, scalability, performance, and costs”
“a new browser that runs entirely on top of Workers”
“Kitesurf started as many other great ideas have started at Cloudflare”
Supabase Pipelines streams Postgres changes to BigQuery in near real time via managed CDC; Unified Logs consolidates all service observability into one queryable interface.
Native CDC and centralized logging reduce operational toil for teams running production Postgres backends. Grafana Cloud integration available on free tier eliminates vendor lock-in for observability.
Pipelines replaces custom Debezium or trigger-based CDC setups on paid plans; Unified Logs replaces log aggregation workarounds. Both are ready now (public alpha and open beta). Grafana integration is one-click and includes free tier, immediate value for monitoring.
“Supabase Pipelines streams Postgres changes to BigQuery in near real time, a managed CDC service you configure in the Dashboard, now available on all paid plans”
“Unified Logs gives you one searchable view across every Supabase service, with live tail, filtering, and a timeline”
“Connect any Supabase project to Grafana Cloud in one click for a pre-built dashboard, alerting, and metrics, available on every plan including Free”
Running `vercel integration add` now pulls agent skills from skills.sh alongside the integration, skipping manual skill configuration.
Agents get immediate provider context without separate setup steps. Reduces the gap between installing a service connector and actually using it in agentic workflows.
Replaces manual skill registration for Vercel-published integrations. Requires CLI update (`npm i -g vercel@latest`) and a skills.sh-compatible provider. Worth testing now if you're building agents on Vercel; no friction if your provider doesn't publish skills yet.
“it now also installs that provider's agent skills from skills.sh”
“This happens automatically for any provider that publishes skills”
“vercel integration discover browses available providers”
“vercel integration categories lists all the Marketplace categories they're grouped into”
vercelagent-skillscliintegrationsautomation
Run coding agents in isolated Vercel Sandboxes
Herdr plugin spawns each terminal agent (Claude Code, Codex, OpenCode) in its own remote Sandbox, eliminating local resource contention and returning changes as Git patches for review.
Developers can parallelize multi-agent workflows without degrading local machine performance, and safely audit all remote edits before applying them. Git-patch returns preserve version control discipline.
Replaces local agent execution with remote sandboxed instances. Requires `vercel link` setup and standard Sandbox billing. Ready now: three agents verified, dry-run safety gate included. Two-invocation flow (dry run, then commit) adds friction but catches mistakes.
“Terminal coding agents like Claude Code, Codex, and OpenCode can now each run in their own isolated Vercel Sandbox, orchestrated from Herdr, a tmux-style manager that runs them side by side in panes”
“Nothing an agent runs or edits touches your machine”
“returns remote changes as a Git patch you apply only when you choose”
“Claude Code (2.1.220), Codex (0.146.0), and OpenCode (1.18.9) are all verified and available by default”