We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Claude subscriptions in third-party harnesses: the 5-hour limit catch
Anthropic renamed Max 20x to Ultra (still $200/month) and Max 5x to Max (still $100). It is also removing the 5-hour limit for verified Ultra accounts, starting with most existing subscribers. Newer accounts keep a temporary limit until they are verified, and weekly limits are unchanged . Theo says Anthropic is vague about what "verified" means. His tip: under settings → usage, if you don't see a current-session bar, your account is verified .
This matters if you use a Claude subscription through anything other than the interactive Claude Code CLI. According to Theo, a user measured T3 Code using about 1.8x more of the 5-hour window per token than the CLI. -p mode has the same problem, and Theo confirmed the numbers himself . Any app using the Agent SDK takes the hit, and he says OpenCode's Claude support uses the Agent SDK too. That leaves the $20 and $100 tiers worse off . The cause, he says, is a cloud code entry point field that is set to SDK-TS. Changing it removes the penalty. He got "semi-solid" permission from Anthropic to report T3 Code as the entry point, so once that release ships, T3 Code should get the same 5-hour and weekly usage as Claude Code directly . He also says he routes his own inference through a CLI proxy, but he won't vouch for the ban risk .
Theo's "council": one review task, several model families
Theo's council skill sends a proposal, plan or change to 3–5 model families for feedback. Because every model gets the exact same "review this change" task, it doubles as a like-for-like eval . He then had Opus and Sol (transcribed as "Soul") blind-judge each other's reviews. Opus picked Opus 66 of 83 times, and Sol picked Opus 54 of 83 times . Median cost per review was $1.65 for Opus versus about $0.36 for Sol. Against what each subscription gives you, he estimates about 205 Opus reviews on Claude's plan versus about 290 Sol reviews on ChatGPT Pro, so the usage gap is much narrower than the price gap . His current split: Claude subscriptions for day-to-day coding, OpenAI subscriptions for computer use and verification, and Grok for Grokbot .
Improving an agent: traces → eval → change
At Runtime 2026, Harrison Chase described a loop you can apply to any harness, including your own coding-agent setup. Run the agent, collect feedback, read the traces, add the failure as an eval data point, propose a change, and check whether it fixes the eval . Harness changes (code, tools, base prompt) start from traces and stay human-reviewed. LangSmith Engine automates the analysis: it runs a sandboxed agent that pulls and clusters traces and suggests fixes, dataset additions and evals . Traces need exact inputs, outputs and tool-call order to be useful for debugging agents or post-training . Evals should cover realistic inputs and sometimes the agent's path, not just the final answer, and they need a simulated environment when the agent is wired to production systems . On fine-tuning, he says SFT on filtered good traces works best for simpler pieces like guardrails or subagents, while RL is better for the core loop but needs an environment and scoring .
LangChain also moved Managed Deep Agents to public beta on LangSmith Cloud . An agent is a folder: agent.py plus instructions.md, skills/ and tools/ . Commands: uvx --from managed-deepagents mda init <name>, uv run mda dev (local server plus Studio traces), uv run mda deploy. There is also a managed-deep-agents skill so a coding agent can do the setup . Each deployment exposes /mcp, so Claude Code can call it as a tool . Coding patterns LangChain reports from beta teams: an issue-tracker webhook that triggers a build-and-verify loop, browser tests in preview environments before human review, and alert triage that opens a fix PR .
Reworking an existing repo instead of starting fresh
Riley Brown gave Codex (GPT-6.1 Sol, extra high) his existing Mac and iOS app repo and the API docs. He told it to read the repo first and reuse the current architecture, then described the new purpose in detail: a screenshot every 30 seconds, judged on-task, off-task or unsure with a confidence score, and nudge rules with thresholds and a cooldown . The build ran for 18 minutes . Next came targeted UI feedback, then a 4x-speed developer mode so the time-based behavior could be tested quickly .
Short notes
- OpenAI Pando: a new Git backend for OpenAI's monorepo that cut push latency by about 80% and serves nearly a petabyte of Git data a day, with no change to how engineers or agents use Git . Greg Brockman says OpenAI is considering offering it to others .
- Weight-space editing: in an MIT study, GPT-6 Astra traced Qwen3.5-0.8B's reasoning loops to two layers and edited 0.016% of its weights. With sampling held fixed, MATH-500 went from 7.9% to 48.1% .
- GitHub Copilot benchmarks: GitHub published per-model performance data for its harness at benchmarks.github.com . Theo had hoped for "more than two existing open source benchmarks" .
- Grounding agents in a point of view: Riley points Claude Code or Codex at specific podcasts through Exa's transcript search before tasks that need cultural context, such as landing pages or scripts .
- Test fan-out small first: Legora's tabular review runs the first 10 rows so you can iterate before scaling to thousands of documents . Its lists primitive triggers agent runs the way you might coordinate coding agents through Linear .
- Automation for unglamorous work: Hex's automated system made 110 fixes in a week, mostly cross-platform compatibility issues the team says it would never have done by hand .
- Evals by side agent: @badlogicgames had a spare agent build a tool plus an eval across models. The eval runs in codemode and spawns subagents in the same durable pi session .
- Open problems in context engineering: LangChain lists async subagents notifying a still-running main agent, deciding what memory should keep, and goal-gaming (e.g., "get rid of all bugs" → delete the codebase) .
- dots model routing: OpenAI's Tibo says dots will route among models including GPT-6 Astra and GPT-6.1 Terra. This is in progress, not launched .
- Riley used Codex to rework an existing macOS/iOS app rather than start from scratch: he supplied the repository and docs, told the agent to inspect the code and reuse its architecture and features, then described the new behavior; the build took about 18 minutes.
- His prompt specified a focus-monitoring loop: capture a computer screenshot every 30 seconds and send it with the task, elapsed time, idle duration, and recent check history to the Decisions API for an on-task/off-task/uncertain choice. He asked for brief detours to be tolerated, repeated high-confidence off-task checks to trigger intervention after at least two minutes, a slower 3–4-minute threshold for 50–80% confidence, and a cooldown between nudges.
- He tested the paired phone and Mac, then gave targeted UI feedback (make the timer larger and simplify the controls); a 4× developer mode sped up testing, and he verified an off-task nudge and session history. Riley estimated eight hours of 30-second screenshot checks would cost roughly 9–10 cents.
- Anthropic's $200/month Ultra tier removes the five-hour cap only for verified accounts; newer unverified accounts retain a temporary limit, while the $20 Pro and $100 Max plans still have five-hour limits.
- Theo reported that T3 Code's Agent SDK integration used its five-hour allowance faster than interactive Claude Code. After receiving tentative permission from Anthropic, he planned to report T3 Code—not SDK-TS—as the entry point, expecting the next release to give it the same five-hour and weekly usage as direct Claude Code. He said his own inference already ran through a CLI proxy, but cautioned that he could not vouch for account-ban risk.
- The sponsored Grapile segment describes code-review agents that run changes in a sandbox, an editable code-and-docs knowledge base that updates with the codebase and is exposed over MCP, and selectable review tiers—from cheaper quick feedback to Apex reviews for large, 20-plus-file PRs.
- Treat a coding agent as a model, harness, and context: the harness handles context, skills, filesystem persistence, and code execution. LangChain’s Deep Agents is an open-source, model-agnostic example of this harness approach.
- Improve the harness by inspecting traces, diagnosing issues, and changing agent code, tools, or the base prompt; keep a human reviewing code changes rather than automatically pushing them to production. LangChain’s Engine approach runs in a sandbox, pulls and clusters traces, uses subagents to analyze them, and proposes fixes, evals, or dataset additions.
- Record exact inputs, outputs, and tool-call sequences, and normalize agent runs into trajectories for downstream models, agents, or post-training. Build evals around realistic user requests and, where relevant, the agent’s path—not just its final answer—and use a simulated environment when testing against production-connected systems is unsafe. A practical improvement loop is to run the agent, inspect feedback and traces, add an eval example, propose a change, and check whether it fixes the eval.
- For model tuning, filtered successful traces can support SFT—through managed offerings or Modal—for simpler tasks such as classification, summarization, and extraction, including guardrails or subagents. RL is a more powerful option for improving the core agent loop, but requires an environment and scoring runs; SFT can establish a baseline before RL.
Sanfilippo describes the paper’s results as preliminary . Astra reportedly identified which game a neural network implemented from its weights without tools or a forward pass , and repaired a buggy network by changing 3.5% of its weights, also without tools or a forward pass . In a separate test with access to activations, Astra corrected a looping Qwen 3.5 0.8B model .
- Riley used Codex to rework an existing open-source Mac+iOS app repository, instructing it to read the repo and reuse its architecture, then update and test both apps; the iPhone pairs with the Mac app via QR code.
- For focus monitoring, the prompt specifies sending a computer screenshot every 30 seconds with the declared task, elapsed time, idle duration, and recent check history to the Decisions API. It uses a choice among on-task, off-task, and uncertain, with confidence; the nudge policy should allow brief detours and wait for repeated confident off-task checks—at least two minutes near certainty, or three to four minutes at 50–80% confidence—before intervening, with a cooldown.
- Riley sped up testing with a 4× developer mode and demonstrated an off-task nudge and a session summary. He estimated eight hours of checks at a 30-second cadence would cost about 9–10 cents, and described Decisions API output as free.
- Coordinate agents through a task layer: Lora’s configurable lists use triggers and automations to launch agent runs and pass results back; agents can also create and read lists. The team explicitly compares this pattern to coordinating coding agents through Linear or another task-management tool.
- Test fan-out on a small sample first: Lora’s tabular review runs the first 10 rows so users can check whether results make sense and iterate before scaling the same questions across a large document set.
- Evaluate models by task and watch for regressions: Lora tracks performance across legal practice areas and tasks in a heat map; its evaluations split work such as due diligence into red-flag detection and report quality, and model releases can improve some tasks while regressing on others.
- Route models and build the harness around the task: Lora uses different models for different tools and routes work as it is subdivided. Its team sees substantial gains available in the harness and retrieval pipelines before model training, and uses post-training for bounded tasks when it can improve cost and latency; for long-document edits, it breaks work into chunks and orchestrates the edits, changing the scaffolding as models evolve.
Willison used GPT-6 Astra Extra High to update his Census business-AI adoption chart with data published October 22, 2026; the chart data came from the Census Employment Size Class XLSX, and he had previously used GPT-5 data analysis and tried Claude Code Interpreter on the same chart. The results require a comparability caveat: on November 17, 2025, the Census changed its question from AI use in “producing goods or services” to use in “any of its business functions,” making responses before and after the change not directly comparable.
Rauch recommends keeping options open for agents by owning the harness, data, connections, and runtime, and continuing to use local models and open source; he cautions that the largest models may remain out of local reach. Kent C. Dodds noted that kody.codes is open source while linking to Rauch’s post.
A cited MIT study reports a weight-space programming result: GPT-6 Astra traced Qwen3.5-0.8B’s tendency to spend 32K tokens on a math problem without answering to two layers, edited 0.016% of its weights, and raised its MATH-500 score from 7.9% to 48.1% with sampling held fixed.
Greptile advertises a code reviewer that uses Computer Use to open an app and click through it to catch bugs; developers can configure its test environment and define the key user journeys it tests.
A Latent Space episode with MIT’s @a1zhang frames Claude Code, Codex, and Pi as broadly similar and describes Recursive Language Models (RLMs) as using code, context offloading, and recursive subagents to generalize across tasks; it also discusses when one expert may replace a trillion-token brute-force search and an OpenAI experiment involving 10,000 agents and 130B output tokens.
- Theo’s “council” workflow sends a proposal, plan, or change to 3–5 model families for feedback, giving each the same review task for like-for-like comparison. In a blind comparison of Opus and Soul reviewing work from both, reviewers preferred Opus: Opus chose Opus 66 of 83 times, and Soul chose Opus 54 of 83 times; Soul was slightly faster at reviews.
- His personal tool split is Claude for day-to-day coding, OpenAI subscriptions for computer use and verification, and a Grok subscription for Grokbot, which he found surprisingly solid.
- Riley uses Exa’s podcast-transcript search API to have an agent search thousands of episodes and recommend listening, rather than rely on podcast-recommendation apps . Exa says its index covers 3 million-plus podcasts, with transcripts spanning millions of hours and updated daily .
- Reusable agent-context workflow: before a task needing cultural understanding—such as creating a landing page or content script—search a specific podcast or group of podcasts and give that perspective to an agent; Riley names Claude Code, Codex, and Grokbot, and says this improves tone, niche understanding, and accuracy . He says his earlier custom-built solution was unsatisfactory, while this one works well and is fast .
Greptile’s sponsored code-review tool uses Computer Use to open an app and click through it to catch bugs; developers can configure its test environment and specify key user journeys for testing.
Kent C. Dodds pointed developers to Kody.codes, saying it can help “get more done within your token limit”; the post gives no workflow details.
Kent C. Dodds announced an upcoming San Francisco workshop at AI Engineer titled “Closing the Agentic Loop: Getting Agents to Safely Build, Verify, and Ship”; he said applications to attend were still open, but did not share the workshop’s techniques in the post.
Riley recommends having agents such as Claude Code, Codex, or Grokbot search millions of podcast transcripts and use a selected podcast or group as perspective before tasks such as creating a landing page, to ground cultural context and improve tone and niche understanding. He says his earlier custom-built approach was unsatisfactory, while this solution is fast and works well.
Theo said T3 Code added an in-app GitHub-star prompt after users sent five prompts; he reported that daily stars immediately increased 10×, a reminder that star counts can respond sharply to in-product solicitation .
Treg announced an open-source Clay alternative for people search and GTM that can run directly in Claude Code or Codex; it claims to be 85% cheaper than Clay, 25× faster, and the most accurate on its people-search benchmark. Jason Zhou says Treg charges agents per answer with 0% markup on provider rates, no subscription, and $0 for misses—a pricing model intended for agents that retry, branch, or return no results.
-
Managed Deep Agents is in public beta on LangSmith Cloud and runs the open-source Deep Agents harness with a managed production runtime. Define the agent as a project folder with
agent.py,instructions.md, skills, and Python tools;uv run mda devstarts a local Agent Server and opens LangSmith Studio for testing and trace inspection, whileuv run mda deployships it to LangSmith. Addmemory.py,schedules/, orsandbox/for cross-thread memory, scheduled runs, or an isolated filesystem and shell. - A reusable coding workflow is to start an agent from an issue-tracker webhook, have it run a build-and-verify loop, and test fixes in preview environments with a browser before human review; incident-triage agents can investigate alerts in a sandbox and, in some teams, open a pull request with a fix.
-
A deployed agent exposes an
/mcpendpoint that Claude Code can call as a tool, and LangChain provides amanaged-deep-agentsskill to guide coding agents through setup. LangChain reports that Content Studio's code shrank by about 80% after moving to Managed Deep Agents.