Build an Autonomous Browser QA Agent with Playwright: Self-Healing Selectors

Tencent Cloud's TencentDB Agent Memory has passed 20,000 GitHub stars within 90 days of open-sourcing, as of August 13, 2026, and the same release adds Team Memory so shared agents can share conversations, documents, code, and institutional knowledge. We map the agent memory stack, the hybrid vector-plus-KV-plus-graph retrieval design, the governance risks of shared memory, and the token economics.

In the week of August 19, 2026, Cloudflare launched Kitesurf, a browser runtime built for AI agents on the Workers platform, claiming roughly 3-7x less CPU and memory than Chromium while passing more than 235,000 web platform tests. We compare it to headless Chromium and Playwright, price the memory tax on agent farms, and show the Worker API shape for agent loops.

Y Combinator open-sourced QM (MIT license, GitHub org yc-software) as a multiplayer multi-agent harness built for a whole company rather than a single user - it runs YC's own accounting, legal, events, and engineering work, including developing QM itself. Cloud-first with native Slack and web interfaces and model-agnostic across Pi, OpenCode, Codex, and Claude Code. This dispatch builds company-harness, a LangGraph workflow with a task intake router, a role router dispatching to role-appropriate agents on swappable model backends, an approval board for cross-role decisions, unified observability and audit, and a self-improvement loop that files tasks back into the queue.

Tencent Cloud's TencentDB Agent Memory crossed 20,000 GitHub stars in 90 days (August 13, 2026) and shipped Team Memory, extending long-term agent memory from a single user to a whole team - conversations, documents, code, and institutional knowledge shared across agents. This dispatch builds team-mem, a LangGraph workflow with a memory-writer that extracts durable facts into a vector + knowledge-graph + KV pool, a memory-reader that fuses vector similarity, entity linkage, and recency behind per-team namespaces and ACLs, and a conflict-resolution node that supersedes contradictions.

On August 17-18, 2026 Alipay unveiled China's first full-stack agentic-commerce platform, letting merchants convert pages and service workflows into agent-ready Skills and MCP tools consumed through Ah Bao (10,000+ AI services, 5 smartphone brands, 16 automakers) over the AHA protocol suite and ACT 2.0 trust protocol. This dispatch builds commerce-orch, a LangGraph workflow that takes consumer intent, discovers Skills, performs a cross-provider A2A-style handshake, gates everything behind a spend-ceiling trust layer, executes payment and fulfillment, and appends every decision to a hash-chained audit ledger.

An unreleased Anthropic model made significant — not complete — progress on the Riemann Hypothesis by testing 650 ideas across 60 parallel subagents over 31 million tokens, with findings confirmed by in-house mathematicians and formalized in Lean. The run is a blueprint for research agents that don't hallucinate proofs: an idea registry prevents duplicate exploration, and a Lean formalization gate means hypotheses only count if they compile. We diagram the fan-out pipeline, stage by stage, and price the token economics from roughly $12k to $62k depending on routing. This is progress, not a proof, and the article is precise about that.

Nvidia released Nemotron 3.5 Lightning, a 30B-A3B MoE open model that is roughly 4x faster at output and 30% faster at agentic task completion, and open-sourced NeMo Switchyard, a routing library that picks the optimal model per request. Nvidia reports a Switchyard-routed stack cuts completion cost to about a third of running Claude Opus 4.8 alone; partners report 21% lower latency (Boomi), 58% lower cost (Ramp), and 74% lower cost at a 6% accuracy tradeoff (LangChain). Meanwhile Nemotron 4, a 1T+ parameter flagship, is in training with a ~$7B cloud-compute budget through FY2028. We map the family, model the routed unit economics, and show a routing-policy implementation.

On August 13, 2026, OpenAI previewed Ultrafast, a serving tier that streams GPT-5.6 Sol at up to 750 output tokens per second on Cerebras wafer-scale engines, with up to 14x the throughput of the Standard tier. Cerebras-reported benchmarks put it ~11x faster than Claude Fable 5 and ~5x faster than Claude Opus 4.8 Fast, with a 5.6x end-to-end speedup on GDP-Val. We break down the latency-budget math of multi-hop agent loops, the ROI for support and incident-response agents, and the API code for benchmarking the new service tier. All figures are vendor-reported preview data.

Anthropic's unreleased frontier model tested 650 ideas across 60 subagents (31M tokens) on the Riemann Hypothesis — and the confirmed findings were formalized in Lean, making Lean the emerging proof-of-correctness gate for both math and AI code. This dispatch builds lean-mcp, a FastMCP Python server exposing five tools — check_lemma, compile_project, list_theorems, propose_tactic, check_import — with inputSchema, mcpServers config, and a localhost transport secured by OAuth 2.0-scoped registry tokens.