Agent deployments need spend gates, identity chains, and measured local runtime paths

The agent budget RFC treats cost as a pre-call authorization decision. It proposes estimating worst-case spend before each provider request, reserving that amount atomically across run, user, key, and other scopes, then committing actual usage after the call. The design targets loops that resend growing context; the RFC says one run can pass 50K input tokens by step 20.

The useful detail is the failure path. Agents receive budget state through headers and RFC 9457 problem-detail errors, so they can choose a cheaper model, shorten context, or stop cleanly. The RFC cites reported cost incidents, including a $4,200 weekend bill and an $87K monthly team bill, but it does not report latency, savings, false-block, or adoption results.

The authentication article gives a practical rule for AI access: every action should inherit authority from a human user and pass through deterministic authorization. For retrieval-augmented generation (RAG), chunks carry access metadata and retrieval results are filtered before the model sees them. For tools, the article points to Model Context Protocol (MCP) with OAuth 2.1 or ordinary APIs with tokens and gateway controls.

The agent section adds separate identities for sub-agents, short-lived credentials, signed JSON Web Tokens (JWTs), delegation claims, and audit logs. The banking example splits work among four agents and uses a 300-second JWT lifetime. The piece is guidance, not an evaluated security study, so its value is in the concrete access pattern.

Donobu packages AI browser actions inside Playwright tests. A test can call page.ai() with a natural-language instruction, constrain it with schemas and tool allow-lists, and cache generated tool calls beside the spec. The package also records failure artifacts, including screenshots, DOM dumps, GPT summaries, and treatment plans.

The auto-heal path is the main operational claim. With --auto-heal, Donobu can rerun an autonomous repair flow and attach a regenerated fixed-test.ts. The corpus gives no pass rate, cost, latency, or baseline comparison, so the grounded read is narrower: the tool turns AI testing into a traceable workflow with cached actions and failure evidence.

Kortex is the strongest measurement-heavy item in the period. It runs models larger than GPU memory by streaming weights across NVMe, RAM, VRAM, and GPU compute. On one Windows 11 machine with a Radeon RX 7900 XT 20 GB, 32 GB RAM, and two NVMe drives, it reports Llama-3.3-70B Q4_K_M at 1.95 tokens per second.

The comparison target is llama.cpp b9860 Vulkan with 30 of 80 layers offloaded on the same hardware, reported at 0.21 tokens per second. Kortex attributes the gap to keeping all compute on the GPU while streaming weights over PCIe. Its limits are also clear: Windows-only streaming today, no HTTP server or multi-turn REPL, and no tested Linux path yet.