AI Workflows & Agentic Architecture Directory

Cloudflare shipped Kitesurf (week of Aug 19 2026), a browser runtime built for AI agents that runs on Cloudflare Workers using roughly 3-7x less CPU and memory than Chromium while passing more than 235,000 web platform tests — a lightweight alternative to full headless Chromium for agent web automation. This dispatch builds browser-runtime-mcp, a FastMCP Python server exposing ten governed tools — open_page, get_text, click, type_text, screenshot, wait_for, fill_form, extract_table, session_state, close_page — where every call returns a compact accessibility-tree view instead of raw HTML to save tokens, with a concurrency/rate-limit guard, privacy mode (trackers/ads blocked by default), and a per-agent session sandbox.

On August 17-18, 2026, Alipay unveiled China's first full-stack agentic commerce platform in Hangzhou, letting merchants with no AI convert pages and workflows into agent-ready Skills and MCP tools. It connects to Ah Bao, the consumer agent with 10,000+ services across eight categories, and ships a full protocol suite (AHA plus ACT 2.0) for audited, intent-verified agent transactions. We break down the stack, the protocol comparison, the merchant ROI math, and a sample skill manifest.

Reuters reported on August 13, 2026, that Anthropic is in talks to acquire Nvidia-backed Decart AI ahead of a potential listing, with Bloomberg pegging the deal at around $6 billion. Decart brings Lucy (real-time live video editing) and Oasis (simulated environments for robotics and autonomous driving), and the team would join Anthropic's inference and performance organization. We analyze why a model lab buys infrastructure pre-IPO, the vertical-integration cost logic, and what it signals to rivals.

Per press reports, OpenAI is targeting an IPO as early as September 2026, with a public S-1 expected within weeks that would expose ChatGPT's real revenue, margins, and unit economics for the first time. OpenAI has told investors enterprise revenue has overtaken its ChatGPT consumer business. We break down what the S-1 will reveal, the Luna/Terra/Sol tiering context, margin scenarios, and what real numbers mean for every AI company raising money.

On August 13, 2026 OpenAI previewed Ultrafast, a service tier running GPT-5.6 Sol up to 14x faster on Cerebras wafer-scale engines at 750 tokens per second. This dispatch builds fastres, a LangGraph incident-response workflow with a latency-critical fast path, a hot diagnostics cache, a premium-budget guard, and degraded fallback to standard inference. It routes S1/S2 events to the fast tier, verifies root-cause hypotheses, drafts a runbook, gates on human approval, and closes with a post-incident review.

Anthropic reported its unreleased frontier model made significant progress on the Riemann Hypothesis by testing 650 ideas across 60 subagents with 31 million tokens, formalizing the confirmed findings in Lean. This dispatch builds leanverify, a LangGraph formal-verification workflow that decomposes a conjecture, fans out parallel research subagents, gates every promising lemma on a Lean compile, dedupes failed ideas in a shared registry, and publishes a verified-claims ledger.

Bloomberg reported Temporal is in talks for a roughly $500 million round at a valuation of at least $12 billion, more than doubling its February 2026 $5 billion Series D — durable execution has become the agent backbone. This dispatch builds duragent, a LangGraph workflow with checkpoint-and-replay semantics: every node result persists to Redis, crashes resume from the latest checkpoint, side effects dedupe on step IDs, and a saga node reverses partial work on failure.

AWS and Unsloth published four deployment patterns for quantized LLMs across EC2, SageMaker, EKS, and ECS that cut inference memory by roughly 75% and cost by up to 80%. The win comes from INT4/FP8 weights plus KV-cache and instance-class downshift, with Unsloth accelerating the fine-tune-to-GGUF pipeline at 2x speed and 70% less VRAM. We compare the four patterns, model the unit economics of a quantized 8B on g4dn vs g5 vs CPU, and include vLLM, Unsloth, Fargate, and EKS code. Treat the numbers as recent-trend guidance and validate on your own fleet.

On August 11, 2026, OpenAI made Daybreak cyber capabilities available through Amazon Bedrock. Daybreak Blue wraps GPT-5.6 Sol and other frontier models with safeguards for authorized defensive work; Daybreak Red provides purpose-trained models for vulnerability research, exploit validation, and security testing. We map Blue vs Red, diagram an agentic security-triage pipeline with a human approval gate, and lay out the access-tier, audit-log, and sandboxing requirements for running cyber models in an enterprise. Authorized and defensive use only.

Bloomberg reports Temporal is in talks to raise roughly $500 million at a valuation of at least $12 billion, more than doubling its $5 billion February 2026 Series D. The bet is durable execution: workflows that persist across crashes, retry from checkpoints, and survive redeploys — now the backbone of AI agents that run for hours or days. We explain deterministic replay over event-sourced history, price the unit economics of losing an uncheckpointed 6-hour job, compare naive agent loops with durable workflows, and show Temporal Workflow and Activity code with retry policies and idempotent IDs.

On Aug 13 2026 OpenAI previewed Ultrafast — GPT-5.6 Sol at up to 750 output tokens per second (~14x Standard), powered by Cerebras wafer-scale engines (900,000 cores, 44GB SRAM on one slab). This dispatch builds cerebras-fast-mcp, a FastMCP Python server exposing four governed tools — complete_fast, tokens_per_second, route_fast_vs_standard, failover_standard — with a TTFT latency guard, a per-request cost/jitter meter, inputSchema, mcpServers config, and OAuth 2.0 scoped-token security.

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.