Example briefs

Claude Code has crossed from developer curiosity into infrastructure-grade engineering tooling. The dominant signal this month: the practitioner community is converging on 'loop engineering' as the successor paradigm to prompt engineering, Anthropic reportedly launched Claude Code Routines (scheduled/event-driven autonomous runs) as a research preview, and a new Claude model (Sonnet 5) reportedly shipped as the default. Simultaneously, real production deployments are surfacing concrete failure modes — MCP tool gaps causing divergent cost calculations, context loss between sessions, and architectural drift in long-running agent projects. The tooling ecosystem around Claude Code is exploding (memory layers, reliability testers, workflow composers, cost reducers), signaling a maturing but still rough production landscape. Key caveats: Routines and Sonnet 5 details rest on a single non-Anthropic source each and are unconfirmed by official Anthropic communications found in these findings.

Anthropic is navigating its most turbulent 30-day stretch yet: a reported Claude Fable 5/Mythos dual-model suspension under US government export-control orders, a forced global access cutoff that sparked international backlash, an unconfirmed IPO filing, and a widely-cited ~$965B valuation — all while shipping Claude Tag for Slack, rolling out agentic features, and managing mounting user anger over pricing, refunds, and identity verification changes. Many of the most dramatic claims rest on social sources or single secondary outlets; treat hard facts (IPO filing, valuation, government directive terms) as reported-only until confirmed by primary sources.

The AI coding-agent space saw a historic capital concentration burst in the last 30 days: SpaceX reportedly acquired Cursor (Anysphere) for $60B in an all-stock deal, OpenAI announced an agreement to acquire cloud infrastructure startup Ona to power long-running Codex agents, and Cognition (Devin) raised $1B at a $26B valuation. All major deal figures rest on media reports and aggregator sources — not confirmed SEC filings or official disclosures — and should be treated as reported, not settled fact. Practitioners on the ground are focused on harness quality, memory persistence, and audit trails rather than model benchmarks.

AI agents in legal tech have crossed a visible inflection point in the last 30 days: a regulated AI law firm reportedly won its first contested UK court case (The Guardian), Thomson Reuters is reported to have relaunched CoCounsel on Anthropic's Claude Agent SDK, Harvey is reportedly building a custom agentic model, and Gartner (via Global Legal Post) projects in-house legal tech budgets will double by 2028. Simultaneously, courts are sanctioning attorneys for AI-fabricated citations and practitioners are raising urgent alarms about autonomous agents violating client confidentiality. Confidence on several headline claims remains 'reported' — single-source — and key hard facts (Spellbook hire terms, BARBRI acquisition, Harvey model specs) rest solely on social or unverified secondary sources. The opportunity window is real but the liability and governance gap is widening faster than most firms are moving.

Agentic payments has moved from concept to infrastructure build-out in the last 30 days. Visa plugged its network into ChatGPT; Mastercard announced 'Agent Pay for Machines'; Stripe Sessions explicitly framed agents as 'autonomous economic actors responsible for most internet transactions.' The race to own the rails for AI-driven commerce is live — but identity verification, fraud, and chargeback liability for agent-initiated transactions remain unsolved and are the central battleground.

Agentic AI for legacy transformation has crossed from experimentation into active productization. Major platform vendors (AWS, Microsoft, TCS, IBM/ServiceNow, SAP) have shipped or previewed agentic modernization tools in the last 30 days, while practitioners are simultaneously surfacing real reliability and governance friction. The market is moving fast but the gap between vendor claims and production-grade trust remains wide.