Claude Code 2.0 Auto Mode, SKILL.state's 94% Token Cut, and Uber's AI-Authored PR Pipeline
This post is a structured English summary of a Chinese-language zhichai.net forum post covering five AI coding-agent engineering inflection points reported around August 30, 2026 (sources: AGI Hunt AI News Daily, GitHub Changelog, Warp engineering blog, Anthropic Claude Code changelog, Uber Engineering).
Key points
- Claude Code 2.0 makes Auto mode the default: agents evaluate task boundaries and invoke tools without per-action approval, shifting the product from "human reviews, machine executes" to "human supervises, machine self-reviews." New capabilities include cross-session messaging (agents hand results between sessions), multi-agent support (parallel agent instances per project), and self-hosted environments for enterprise-internal deployment.
- Community stress test: a developer reportedly ran 12 Claude Code agents simultaneously on a 16GB M1 Pro MacBook, each driving a functional iOS simulator — enabled by agents self-managing memory and context windows under Auto mode.
- Silent transcript deletion: Claude Code defaults to
cleanupPeriodDays: 30, permanently deleting local transcripts older than 30 days. Workaround: set it to3650(~10 years) in~/.claude/settings.json. - Google SKILL.state: instead of compressing history, agents write a structured current state each step plus a latest observation, then explicitly discard history. Reported result: ~94% reduction in long-session token usage versus feeding full history as context — the LLM's attention budget goes to decisions rather than recall.
- Uber's software factory: over 70% of pull requests are now attributed to AI agents (Claude Code, Codex, and internal agents registered as developer identities in git blame). One team merged 214 PRs last month but humans fully read only ~12; the rest were skimmed via summaries and test results, supported by automated review tools like CodeRabbit (flagging NPE risks, coverage gaps in seconds).
- Warp's dual-skill split for self-improving code review addresses two failure modes: models not knowing project norms, and forgetting prior findings. Skill 1 (review) examines the current diff; Skill 2 (consume) drives fixes from findings. Their Atomic paradigm models tasks as recursive state machines with an explicit verifier (tests, coverage, style) rather than agent self-declared completion.
- Will Claude Code 3.0 push Auto default to long-running Auto, driving human review toward 0/214?
- Will agent-hires-agent economies (Hyr/Mandate) trigger regulatory licensing as micro-financial institutions?
- Will "Human Reserved" become law or industry norms mandating human-reviewed PR ratios?
- AGI Hunt AI News Daily 2026-08-30
- Warp engineering blog (dual-skill review methodology)
- GitHub Changelog: Claude Code v2.1.251
- OpenAI Codex CLI v0.151.0 release notes (MCP result hooks)
- Anthropic Claude Code changelog
- Uber Engineering software-factory article
- The Atlantic interview with Bill Gates (Human Reserved proposal)
Long-horizon agent approaches
| Approach | Mechanism | Reported result | |---|---|---| | Google SKILL.state | Externalize state to structured files, discard history | ~94% token reduction | | PILOT | Continuous in-run reflection, updating policy for long-chain decay | Performance recovery on long-horizon benchmarks | | LongHorizon-Harness | State kept outside growing context; only independently verified facts; Manage-Execute-Audit loop | ~3x improvement on OSWorld | | Proteus | Agent rewrites its own harness source (tools, prompts, execution flow) in Observe → Purpose → Act → Reflect cycles | Self-evolving harness | | JIT-Agent | On-the-fly synthesis of scaffolding (memory, planning, tool orchestration) | Weaker models + scaffolding can beat stronger models |
Shared thesis: long-horizon agents fail not from insufficient context length but from not knowing what to remember, discard, and in what structure — agents should query external state rather than carry everything in-context.
Agentic economy: Hyr and Mandate
Hyr (hyr.tech), founded by former OpenAI researcher Charlie Snell, launched a marketplace where users pay agents and agents hire other agents against USD budgets. Its Mandate funding layer (Rust daemon + encrypted double-entry ledger) lets agents earn, hold, move, and spend USD across Stripe, banks, and wallets under a single account. Related ecosystem stats disclosed the same day: Vercel MCP tool calls +564% over three months; 1M+ skill installs via npx skills add; workweave/router decision latency <50ms with 40–70% cost reduction.
Costs and open questions
1. Data retention: vendors are deciding retention policy in defaults (30-day silent deletion) that should belong to users. 2. Structural over-trust: when review tooling makes default-approval rational, a single missed bug class flows through ~70% of PRs automatically — not fixable by adding reviewers. 3. Human position: citing a Gates interview in The Atlantic proposing legally designated "Human Reserved" work, the author asks where humans remain in production pipelines.