KDCube
← AI Industry News

DeepSeek's Unprompted Coding Agent & OpenAI's Settable Compaction

DeepSeek open-sourced Super-Agent, an MIT-licensed prototype that watches a team's repo and chats and fixes code without being summoned — a direct swipe at Claude Code. Meanwhile OpenAI's SDK 3.42.0 makes context compaction and credentials settable per call, and Google's Vertex AI Extensions shutdown clock tic...

At A Glance issue scan

  • signals03
  • watch items03
  • sources06

Highlights

  • DeepSeek open-sourced Super-Agent (MIT license), a coding-agent prototype that watches a team's GitHub repo and chat and jumps in with fixes without being summoned — an explicit shot at Claude Code. (ShiftDelete)
  • OpenAI's Python SDK 3.42.0 (and Node v7.46.0) moved context compaction, session review, and experimental agent time awareness from opaque server defaults into client-settable call options. (AIWeekly)
  • OpenAI also shipped per-request callable API keys and org external key management in the same release — credential scoping that matters for multi-agent fleets. (openai-python releases)
  • Google's Vertex AI Extensions shut-off clock is now live: the service is deprecated and turns off after Nov 26, 2026, with migration to the Gemini Enterprise Agent Platform required. (Google Cloud)

Key Signals ranked scan

  1. DeepSeek ships a self-prompting coding agent

    announced Oct 31 WeChat post; launched Nov 1, 2026

    DeepSeek's Harness team published Super-Agent as a research prototype under the MIT license — an agent that proactively monitors developer group chats and the repo, then proposes fixes without a per-task prompt. DeepSeek labels it "not an official product" and "not production-ready," but the direction — an always-on teammate rather than a summoned assistant — is the newsworthy part, and it aims squarely at Anthropic's Claude Code. (ShiftDelete, Bloomberg)

    agents

  2. OpenAI makes compaction a knob, not a mystery

    Oct 31, 2026

    SDK 3.42.0 lets builders set summarization behavior per call across beta agents, sessions, and scheduled runs, adds session review/tool settings, and an experimental time-awareness flag. Operator catch: background compaction is on by default and can increase token usage, so the win in context hygiene has a billing tail to watch. (AIWeekly, openai-python releases)

    model-ops

  3. Google's Vertex-to-Agent-Platform migration gets a hard deadline

    as of Nov 2, 2026

    All Vertex AI services are now delivered through the Gemini Enterprise Agent Platform, and Vertex AI Extensions shuts down after Nov 26. The platform's TPU 8i inference silicon is pitched at ~80% better performance per dollar for high-concurrency agent workloads — the cost lever Google wants teams to migrate toward. (Google Cloud, datacenter.news)

    platforms

Market Heat topic map

  • Agentscoding
  • Costtokens
  • Opscompaction
  • Govcredentials
  • MCPtools
  • TPUinference
  • OSSruntimes
  • Migrdeadline

Why It Matters / What To Watch

  1. "Always-on" agents are the new coding-agent frontier.
    • Treat Super-Agent as a signal, not a dependency — it is MIT-licensed and open, but DeepSeek warns it differs from its internal tooling and isn't production-hardened. (ShiftDelete)
    • Watch whether proactive, chat-watching agents force governance questions (who approved this fix? what did it read?) that today's summon-on-demand assistants sidestep. (Bloomberg)
  2. Compaction and credentials are now your settings to own.
    • Audit the new default: background compaction on can quietly raise token spend — set it explicitly per run and meter the delta before rolling to fleets. (AIWeekly)
    • Adopt per-request callable keys to scope agent credentials rather than sharing one long-lived key across a swarm. (openai-python releases)
  3. Clock-watch the Vertex Extensions cutover.
    • If you still call Vertex AI Extensions, schedule migration before Nov 26 or expect disruption. (Google Cloud)
    • Re-benchmark inference cost on TPU 8i before assuming the perf-per-dollar claim holds for your concurrency profile. (datacenter.news)

Quick Links sources