06-reference

alphasignal chatgpt work agent claude reflect swe17

2026-07-10·reference·source: AlphaSignal·by AlphaSignal
AImachine-learningcurationdaily-digest

"ChatGPT Work Agent 🏗️, Claude Reflect Dashboard 🧘, SWE-1.7 at $1.97/ta" — AlphaSignal

Why this is in the vault

Three items in this issue land directly in RDCO's operating stack: Claude Code's /advisor command (lets any model review your code — unknown-to-many capability), SWE-1.7 pricing at $1.97/task (cost anchor for agentic coding decisions), and Claude Reflect (Anthropic's new usage dashboard with quiet hours). The GPT-5.6 Work agent is tier-1 competitive intelligence on OpenAI's workflow-ownership bet vs. Anthropic's.

⚠️ Sponsorship

Three paid sponsor blocks in this issue:

Signal #2 ("Coding agents can't fix bugs they can't see. Wire Sentry MCP…") is a paid placement, not editorial.

Issue contents

Top News

1. OpenAI ships ChatGPT Work (GPT-5.6 agent) A new ChatGPT mode that runs full workflows for hours autonomously. Connects to Google Drive, Slack, Salesforce and other apps. Outputs finished docs, slide decks, spreadsheets, and shareable web apps. GPT-5.6 comes in three tiers: Sol (top), Terra (mid), Luna (budget). Codex merged into ChatGPT desktop (Mac/Windows). Live now for Pro/Enterprise/Edu; rolling to Plus/Business soon.

2. Anthropic ships Claude Reflect A usage dashboard inside Claude showing most active day, peak hour, total conversations, topic breakdown, and workflow tips. Includes controls for quiet hours and break nudges after a chosen usage amount. Covers 1/3/6/12 month lookback windows. Skips incognito chats and health-integration data. Access via Settings > Reflect (requires Memory on, web or desktop).

3. Cognition releases SWE-1.7 Coding model built on Kimi K2.7 + reinforcement learning. Scores 42.3% on FrontierCode — near Claude Opus 4.8 and GPT-5.5 level — at $1.97/task. Runs at 1000 tokens/second via Cerebras. Trained across 4 datacenters on 3 continents. Investigates more files before editing (touches more files as tradeoff). Available in Devin (Web, Desktop, CLI).

Signals (editorial)

  1. Open-source voice-cloning tool that pipes into Claude or Cursor (6,515 likes)
  2. (Sentry sponsored) Sentry MCP for production debugging context in coding agents
  3. Zhipu AI SAO — RL training trick beats GRPO on coding and math benchmarks (Zhipu's SAO makes agentic RL training leaner — entropy collapse and numerical drift fixes)
  4. Open-source benchmark testing 30+ robot models across sim and real-world tasks
  5. Cohere Labs releases Aya Vision, 8B open-weights, 23 languages (144k downloads)
  6. Claude Code's hidden /advisor command lets any model review your code (3,694 likes)

Mapping against Ray Data Co

Claude Code /advisor command — Signal #6 is directly actionable. The /advisor command in Claude Code lets you route a code review to any model (not just the one driving the session). RDCO runs Claude Code as the always-on COO agent; this unlocks model-swapping for review passes without leaving the Claude Code environment. Worth testing in the RDCO agent loop.

SWE-1.7 at $1.97/task — Provides a concrete cost anchor for agentic coding benchmarking. At 42.3% FrontierCode and near Opus 4.8 level, this is the cheapest frontier-adjacent coding agent available. Relevant when pricing out the cost of automated coding tasks in the RDCO stack vs. Claude Code's token-spend model.

ChatGPT Work (GPT-5.6) — Competitive read: OpenAI is betting on deep workflow ownership (hours-long autonomous execution, multi-app integrations). Anthropic's counter is Reflect/quiet hours — a very different surface-area bet. The divergence matters for RDCO's choice of agent substrate over the next 6 months.

Zhipu SAO RL research — Lower-priority but relevant to understanding why next-gen coding models (like SWE-1.7) are getting better fast. Entropy collapse fixes = more stable long-RL runs = cheaper capability gains in coding agents.

Related