Your stack
You are
Mostly building
Your edition
20 items
Relevant to you · Agents & coding tools
OpenAI
Agents SDK now available in TypeScript with sandbox support
5 May · OpenAI released a TypeScript version of their Agents SDK with sandboxed execution, letting you run agents in isolated environments safely.
Anthropic
Seed managed agent sessions with initial events to start work immediately
21 Jul · You can now bootstrap a managed agent session with context and goals in one call, eliminating a separate send-events request.
Webhooks now track environment and memory store lifecycle events
21 Jul · Managed agents emit events when their environment or memory store changes, so you can trigger downstream work without continuous polling.
Thread-level event streams now support delta previews
21 Jul · Subagent output now streams token-by-token at the thread level, matching session-level streaming for consistent real-time visibility.
Set effort levels on managed agents to control compute allocation
21 Jul · Managed agents now let you tune how hard Claude thinks on each task, balancing speed against answer quality without changing the model.
Claude Code
Subagents can now nest up to depth 3 instead of 1
24 Jul · Agents can now spawn sub-agents that spawn their own sub-agents (up to 3 levels deep), enabling more complex task decomposition. Revert with an env var if needed.
DirectoryAdded hook fires when working directory is registered mid-session
24 Jul · Hook fires whenever a new working directory is registered mid-session, letting you run custom setup or validation logic for newly added repos.
Claude Opus 5 becomes default model with 1M context
24 Jul · Claude Opus 5 replaces 4.7 as the default model, offering 1M context window and faster inference at new pricing. Existing sessions can switch via `/model`.
Multi-line paste no longer collapses to single line
22 Jul · Pasting multiple lines of code was turning them into one line with stray `j` characters in some terminals; that's fixed.
Vercel
DeepSeek V4 Flash weights updated with stronger agentic performance
31 Jul · DeepSeek V4 Flash received a significant model update improving agentic reasoning capabilities; requests automatically use new weights without code changes or model ID updates.
Run multiple agents isolated in one Sandbox
30 Jul · Agents can now coexist in a single Sandbox with filesystem isolation—each gets its own user context, so they can't interfere with each other's files or state.
Install agent integrations from the CLI with registry support
29 Jul · You can now discover, preview, and install integrations for your agents via CLI, with support for custom registries and third-party sources alongside the official catalog.
Agents can cancel in-progress responses or reset conversations
27 Jul · Agents can now interrupt themselves with ctx.cancel() to handle corrections, or ctx.reset() to completely restart a conversation—useful for recovering from mistakes or changing direction mid-stream.
Also waiting
OpenAI
Fast mode replaces Priority Processing for GPT-5.6 Sol
30 Jul · OpenAI replaced its Priority Processing with Fast mode, delivering faster inference for Sol at higher cost with automatic backward compatibility.
GPT-5.6 Luna and Terra pricing drops
30 Jul · OpenAI reduced inference costs for two GPT-5.6 variants, making Luna especially attractive for cost-conscious applications.
OpenAI Terraform provider for API platform infrastructure
28 Jul · OpenAI released an official Terraform provider to codify API platform configuration—projects, access control, and rate limits—alongside your infrastructure.
Vercel AI SDK
generateText warns on streaming-only timeout settings
31 Jul · If you pass streaming-specific timeouts to generateText (which doesn't stream), you'll now get a warning to help catch config mistakes.
streamText and ToolLoopAgent now support experimental tool callers
31 Jul · Tool callers—a new experimental pattern for invoking tools—are now supported in streaming contexts and agent loops, expanding how you can structure tool use.
Telemetry now correctly attributes calls to their actual response model
31 Jul · When you call a model that gets routed or remapped to a different model internally, telemetry now credits the actual model that responded instead of the one you asked for.
Route tool calls through experimental_toolCaller in code mode
30 Jul · You can now intercept and route tool calls in generateText using an experimental handler, useful for custom tool execution pipelines.
200 more waiting. Each tool contributes four items an edition, use-case matches first — the rest come in a later edition rather than being dropped.
Built from entries already ingested and shaped by the same code that builds the email. Your real editions also skip anything you have already been sent.
How it works
Thirty seconds now, an hour a week back.
Pick your tools
Models, coding agents, frameworks, auth, data, infra, UI — check the ones you actually ship with.
Say who you are
One technical level, one use case. Twelve lenses decide how every item gets written for you.
Read one email
Every week, only the changes that touch your stack — already summarized, already sorted.