The industry is shipping production controls and protocol updates that assume agents are already mission-critical infrastructure. This week's reads cover runtime governance layers, stateless protocol upgrades, token-optimized frameworks, a new cost-efficient frontier model, and harness architectures built to outlive the models they wrap.
LangSmith LLM Gateway: runtime controls for production agents
LangSmith's gateway acts as a policy enforcement point that blocks requests before they violate spend caps, rate limits, or data policies rather than alerting after the damage is done. The key shift is centralized governance: you set budgets and redaction rules at the organization or API key level, and the gateway routes calls to multiple providers while enforcing consistent controls. This matters for teams running agents in CI/CD where a single runaway job can burn through monthly budgets in hours.
// Gateway enforces limits before the call reaches the provider
const response = await gateway.chat({
model: "gpt-5.6-luna",
messages: [...],
// Spend cap enforced at request time, not post-hoc
metadata: { workspace: "ci-agents", user: "github-actions" }
});How AgentCore Gateway supports the MCP 2026-07-28 spec
The Model Context Protocol just went stateless, moving from WebSocket-based sessions to standard HTTP with header-based routing. AWS's implementation supports multi-round-trip requests (server-to-client calls), per-request version negotiation, and a governed extensions system, all while keeping existing clients working unchanged. This is significant because stateless protocols scale on commodity infrastructure and integrate cleanly with API gateways, load balancers, and existing observability tooling without special session management.
Deep Agents v0.7
LangChain stripped out 65% of base input tokens by making scaffolding opt-in, trimming tool descriptions, and removing default middleware that newer models don't need. The validation is what matters here: performance held steady across benchmarks while GPT-5.6-Luna token usage dropped 34%. If you're running agents at scale, this translates directly to lower per-run costs without re-architecting your workflows. The release also exposes full middleware control, so you can swap built-in components like SummarizationMiddleware with custom logic.
Introducing Claude Opus 5
Opus 5 delivers near-frontier intelligence at half the cost of Fable 5, with measurable gains on agentic tasks like self-verification and long-horizon consistency. Anthropic emphasizes alignment improvements and lower deceptive behavior rates, though they intentionally kept cybersecurity exploitation performance below Mythos 5. For teams evaluating model swaps, the combination of cost reduction and stronger task persistence makes this a natural candidate for agent workflows that currently burn budget on more expensive models.
Building an agentic harness that outlasts the model
Shopify's Dispatch harness scanned 80+ apps in six weeks, generating 300+ validated vulnerabilities worth an estimated $400k+ in bug bounty value at $50-$300 per scan. The architecture—partitioning, test oracles, cross-model verification, deterministic workflows—is the durable asset, not the model. This aligns with the broader pattern: production agent systems need governance, validation, and control planes that survive model churn and token price swings.
The through-line this week is infrastructure maturity. Teams are building control planes, protocol layers, and harness architectures designed to outlast individual models and absorb cost fluctuations without breaking production workflows.