We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
The coding-agent bottleneck is now verification infrastructure, not code generation. Anthropic reports that Claude authors 80% of its code and engineers ship 8× as much per quarter; tests grew 10× and CI jobs 25× in six months, forcing a redesign of test-impact analysis from a single stateful process into stateless listeners plus a journal/consumer path. Addy Osmani’s brownfield rules and LangChain’s paid-media agent point to the same operating model: autonomy follows blast radius, deterministic code and source boundaries, permissions, and post-change verification—not model confidence.
⚡ TRY THIS
Zone the repository before granting autonomy — Addy Osmani. Have a person map green areas with good tests and isolation, yellow areas with mixed quality, and red areas such as auth, billing, and permissions. Green can run a tight agent loop; yellow requires characterization tests; red requires human pairing. For yellow/red work, run a read-only exploration pass that leaves a short memo citing files, owners, history, tests, and production signals; plan in a clean context; then lock today’s behavior with tests written by a separate pass or person before implementation.
Design the verification path for the agent load you actually want. Anthropic’s test-impact service uses a listener to record every CI result and a selector to choose relevant tests; its redesign moved state into an in-memory journal, let stateless workers append and exit, and used a separate consumer to roll results into per-test history. Anthropic says the distributed version was easier to scale and profile, took one engineer three weeks, and recommends assuming 25× load within two quarters.
Let code own consistency; let the model own judgment. Keep the system prompt as a navigation map, progressively disclose skills, put company-specific context in a wiki, and move date alignment, calculations, source-of-truth rules, and hard safeguards into deterministic code. For large tool catalogs, expose
search → read → runinstead of loading every schema up front. LangChain reports reducing the first turn from about 38,000 to 12,000 tokens at 4× lower cost, while moving calculations into Python made an early reporting workflow 40× cheaper and 13× faster.Earn parallelism with an eval and an isolation boundary. In @businessbarista’s summary of @Vtrivedy10’s LangChain eval masterclass, define a checkable task and verifier, run it in a safe environment rather than production, trace every tool call, and have a second agent mine failures for new evals. Then give each subagent its own report location and completion state, restrict its tools to what the job needs, and parallelize only after one unit has a dependable judge, recovery path, and review format.
📡 WHAT SHIPPED
Sourcegraph Agentic Batch Changes is available to all Sourcegraph Cloud customers. Describe a change once, run it across 10 or 10,000 repositories, let it adapt to repository differences and CI failures, and pay per changeset merged; an unmerged PR costs nothing.
LangChain’s paid-media-agent is open source. The reference implementation includes ad-platform tools, paid-media skills, a sample wiki, reporting, and approval workflows, with deployment to Slack through Managed Deep Agents.
LangSmith LLM Gateway adds model-scoped access control. Lock an API key to permitted models and block every other call automatically; LangChain demonstrates an Opus 5-only key being rejected when used for Sonnet 5.
T3 Code removed the one-thread/one-PR assumption. A development thread can now link to multiple PRs, and the product added GitHub Stacks support.
Deep Agents changed its file-reading format. LangChain’s OSS team reports internal evals showing 15% fewer
edit_fileerrors and 10% lower input-token usage—a useful harness signal, though not a reproducible benchmark from the announcement alone.Rex exposes the terminal as an orchestration surface. In Mitchell Hashimoto’s demo of the Superlogical multiplexer CLI, agents can create sessions, run commands, rearrange splits, change focus, simulate input, wait on scripts, subscribe to streaming events, and retrieve running-process details as JSON.
🎬 GO DEEPER
- Theo’s mobile-port video: use the simulator as the verifier. The useful workflow is concrete: give the agent the simulator, source code, and a clear existing implementation; describe discrepancies when it misses. Theo reports that his SwiftUI T3 port was built 95% with Soul and 5–10% with Astra in one long-lived thread, with roughly 1,000 people on the public TestFlight.
- LangChain’s Managed Deep Agent Slack walkthrough: deploy, trace, then constrain triggers. Adding Slack to an existing project is
slack init→ deploy → authorize; requests and full traces appear in LangSmith. The default is mention-triggered operation; “all messages” and bot-to-bot triggers exist, but all-message mode belongs in a dedicated channel because it will answer irrelevant chatter too.
Richard Socher on reward engineering, 00:49:49–00:51:29. His “make 100 lines of code faster” example is a clean evaluator warning: an agent can move the stopwatch’s end marker to the beginning and claim an instant speedup. For goals that cannot be directly verified, Socher and the hosts point to rubrics and model-judge criteria as the verification layer.
Repo to study: langchain-ai/paid-media-agent. Read the implementation for the separation of sandbox, skills, wiki, live tools, deterministic code, subagent state, permissions, approval cards, and post-write checks—not just the reported marketing results.
Editorial take: The durable agent loop is no longer “prompt → diff”; it is “map → constrain → verify → merge,” with parallelism earned only after the system can explain and recover from failure.