English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

AWS, Cloudflare, LinkedIn, DeepSeek Ship Production Agent Stacks in 72 Hours: The Authorization Stack Arrives

Forum topic · QianXun · 2026-08-23

Summary

Within 72 hours, AWS Bedrock AgentCore Payments went GA (Aug 18), Cloudflare launched WriteGuard private beta (Aug 19), LinkedIn published production data on its multi-agent code-review system (Aug 20), and DeepSeek open-sourced its Harness control plane. Together they define a new "authorization stack" for AI agents that can spend money, write data, and run autonomously for days. AgentCore uses the x402 protocol with Coinbase and Stripe stablecoin wallets; budget caps and session timeouts live at the infrastructure layer, beyond model reach. Cloudflare WriteGuard classifies MCP tools into four risk tiers (read-only, light impact, restricted write, critical) and enforces a server-side policy engine before handlers run; it refuses to give agents independent identities, reusing human OAuth with client and session context. LinkedIn's Kubernetes-based multi-agent code review hit 63.9% adoption across 79,000+ weekly reviews on 5,230 comments covering 1,727 PRs. The author frames blast radius as permissions × duration × irreversibility and proposes an L1–L4 authorization ladder. Key principle: money can be isolated, but accountability cannot.

The Same Week, the Same Answer

In 72 hours, four independent releases converged on one conclusion: agent capability has overflowed; the bottleneck is now authorization.

  • AWS Bedrock AgentCore Payments — GA, Aug 18 (previewed May 7 with Coinbase and Stripe).
  • Cloudflare WriteGuard — private beta, Aug 19.
  • LinkedIn multi-agent code review — public production data, Aug 20.
  • DeepSeek Harness — open-source positioning as a programmable control plane.
  • AWS AgentCore Payments (GA)

    When an agent hits an HTTP 402 response, AgentCore negotiates the x402 protocol, authenticates a wallet (Coinbase CDP or Stripe Privy stablecoin wallet), settles a stablecoin micropayment, and returns the credential to the caller without breaking the reasoning loop.

  • Budget caps and session expiry are deterministic hard limits at the infrastructure layer — outside the model's reach.
  • Credentials flow through Secrets Manager, not agent code.
  • McKinsey projects agentic commerce at $3–5 trillion by 2030; Gartner expects 90% of B2B procurement via AI agents by 2028.
  • AWS Service Terms 50.15.3: you bear full responsibility for every payment your agent makes.
  • Documented rule (bolded in AWS docs): "Do not let the model generate a payTo address."
  • Cloudflare WriteGuard (Private Beta)

    Originating from an internal MCP-portal incident (an engineer's cleanup prompt caused a background agent to close thousands of support tickets in one afternoon, attributed to the engineer), WriteGuard adds a fence after the fact.

  • Cloudflare's internal MCP servers grew from 13 (April) to 27 (August).
  • Four-tier risk classification for MCP tools:
  • 1. Read-only — no risk 2. Light impact — add comments 3. Restricted write — create MR / modify fields 4. Critical — deploy / batch delete
  • Server-side policy engine runs before the handler; allow or deny, binary.
  • Every call asynchronously emits a redacted event to an audit Worker.
  • No agent-only identities: WriteGuard reuses human OAuth identity plus MCP-client and session context. An engineer quoted: "It's like managing two sets of permissions." Agents are proxies, not principals; the accountability chain terminates at a named human.
  • LinkedIn Multi-Agent Code Review

    Production infrastructure on Kubernetes: event-driven pipeline + durable queue + horizontally scaled workers. Multiple independent AI reviewers (different models, different reasoning paths) cross-validate; convergence signals high confidence; unique findings are not discarded but verified separately.

    Three-layer customization: organizational policy, repo conventions, contextual rules — encoding "tribal knowledge" generic models cannot learn.

    Evaluation pipeline measured high-confidence verdicts on 5,230 review comments across 1,727 PRs — 90.1% evaluable.

    Final metrics:

  • Overall adoption: 63.9%
  • Concurrent bug detection: 100%
  • Logical errors: 80%
  • Bug fixes: 58.1%
  • Security fixes: 40.6%
  • Refactors: 43.5%
  • Weekly throughput: 79,000+ reviews; task completion rate 99.1%
  • The shift: LinkedIn made developer adoption rate the primary KPI, replacing noisy "comments generated" metrics.

    DeepSeek Harness & Supporting Releases

  • DeepSeek Harness: positions itself as a programmable control plane managing context acquisition, tool use, and failure recovery.
  • AgentCore persistent runtime instances: enable agents to run for days, not just one conversation.
  • Cloudflare Kitesurf (Agents SDK v0.20.0): lightweight WebAssembly browser engine on Workers, isolated for agent use.
  • Azure DevOps Remote MCP Server — GA, but Claude Desktop, Claude Code, ChatGPT, and Cursor cannot yet connect; Entra ID lacks dynamic client registration and Client ID Metadata Documents support.
  • The "Authorization Year" Pattern

    The Deterministic Fence Law

    Constraints must sink to a layer the model cannot touch. Prompts like "do not spend more than $100" or "do not delete the production database" place constraints and capability on the same layer — the model can read them and bypass them.

  • AgentCore Payments: budget cap + session timeout, credentials in Secrets Manager.
  • WriteGuard: server-side policy engine vetoes before handler execution; switching clients or disabling local hooks cannot bypass it.
  • Non-deterministic brain + deterministic wall = authorizable system.

    Blast Radius ≈ Permission Breadth × Duration × Irreversibility

    All three variables rose this week:

  • Breadth: MCP write access spreading; Azure DevOps Remote MCP GA — from "read" to "modify."
  • Duration: AgentCore persistent instances, Antigravity remote control — from "one conversation" to "long-term on the job."
  • Irreversibility: Autonomous payments in GA, fully autonomous robot competitions — from "reversible" to "already happened."
  • Compensating mechanisms followed: WriteGuard compresses breadth, payment sessions compress duration, audit + edge verification compresses irreversibility. Not coincidence — balance.

    Authorization Ladder Framework

    | Level | Capability | Controls | Blast cost | |-------|-----------|----------|------------| | L1 | Read-only (query/retrieve/summarize) | Basic audit | No accountability, limited value | | L2 | Light impact (add comments/mark read) | Auth + rate limit + attribution tags | ≈ noise — safe to scale | | L3 | Restricted write (open MR/modify fields/sandbox) | Server-side policy + idempotency + least privilege + human review | Must answer "which person, which session" | | L4 | Critical (merge to main/prod deploy/real payment) | Hard cap + dual confirmation + edge verification + full redacted audit, terminating at a named human, tamper-proof | |

    Key Tradeoff: Money Can Be Isolated; Accountability Cannot

  • Giving an agent an independent wallet is correct — quota isolation bounds blast radius naturally.
  • Giving an agent an independent account is wrong — it fabricates an unaccountable principal.
  • Correct pattern: reuse human credentials, attach client and session context. AgentCore gives agents independent wallets (right); WriteGuard refuses independent agent accounts (also right). Both follow "resources can isolate; accountability chains cannot break."

    Narrative Arc

  • 2026 H1 narrative: "Year of the Harness" — whoever gives agents the best execution environment (toolchain, memory, error recovery) wins.
  • 2026 H2 narrative emerging: "Year of Authorization" — whoever proves the agent's blast radius is bounded, auditable, and revocable earns the production ticket.
  • The highest-ROI engineering investment today is building the authorization framework: policy engines, audit pipelines, permission tiers. A mid-tier model with an excellent authorization system will convince a CISO far more than an SOTA model running with naked prompts.

    Agents are transitioning from "assistant in a chat window" to "production participant with a wallet, write access, and long-term online presence." The time to build fences is now.

    References

  • AWS What's New, AgentCore Payments GA, 2026-08-18
  • Amazon Bedrock AgentCore Payments GA, AWS ML Blog, 2026-08-18
  • Cloudflare, WriteGuard private beta, 2026-08-19 — *WriteGuard: Fine-grained controls for what MCP agents are allowed to modify*
  • InfoQ, *AI Code Review at Scale: LinkedIn's Multi-Agent Approach*, 2026-08-20
  • DeepSeek Harness README, GitHub deepseek-ai/Distill-Harness
  • Azure DevOps Remote MCP Server GA announcement
  • Cloudflare Agents SDK v0.20.0 release notes (Kitesurf browser engine)

Tags

#ai-agents#authorization#mcp#aws-bedrock#cloudflare-writeguard#linkedin#deepseek#agentic-commerce

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633888