Bedrock Agents Classic Freezes — CLI Data + AgentCore Define the Stack

Yesterday we covered Claude CTF breakouts, ARC harness settings, and Korea’s sovereign AI capital push. Today the tape flips from eval cages to where you actually run agents in production — and whether the CLI agents your org is buying show up as merged work, not seat licenses.

Three threads: AWS puts Bedrock Agents Classic into maintenance mode (closed to new customers as of July 30, model catalog frozen); Microsoft publishes org-scale evidence that CLI coding agents (Claude Code + Copilot CLI) lift merged PRs ~24% when they stick; Okta signs to acquire Permiso for human / non-human / agentic identity threat detection.

This is a platform migration + rollout math + identity fabric day.

Neon terminal workspace with abstract multi-agent workflow nodes on a dark cyberpunk background

AWS: Classic Is Done for New Builds

Amazon Bedrock Agents launched November 2023. As of the maintenance-mode docs, it is now Amazon Bedrock Agents Classic:

  • No new customers starting July 30, 2026
  • Model catalog frozen on that date — new foundation models land on AgentCore, not Classic
  • Existing Classic accounts keep running; no announced end-of-life
  • Bedrock models, Knowledge Bases, and Guardrails themselves are not deprecated
  • The two gated calls for non-allowlisted accounts: CreateAgent and InvokeInlineAgent

The recommended landing zone is Amazon Bedrock AgentCore: managed harness (declare model, tools, instructions) plus a code-defined path for custom loops. Gateway turns action groups into MCP tools; memory, identity, observability, and code interpreter sit as first-class services instead of bolted-on agent config.

# Mental model (steal this)
classic:
  onboarding: closed_after_2026-07-30
  model_catalog: frozen
  eol: none_announced
agentcore:
  harness: config_based  # model + tools + instructions
  code_path: any_framework  # Strands, LangChain, OpenAI Agents, Claude SDK...
  tools: gateway_mcp
  memory: short_and_long
  identity: first_class
  observability: end_to_end_trace

Builder takeaway: if you greenfield on AWS this week, do not start a Classic agent. Import Classic configs via the AgentCore CLI if you already have them; use the managed harness unless you need stage-specific prompt overrides or full multi-agent routing that the harness does not yet express. Knowledge Bases stay; you reconnect them through the gateway.

If you are standing up a home lab that mirrors “runtime + gateway + memory,” a small always-on NUC / mini PC plus a spare NVMe for agent traces beats re-running failed sessions because you never kept the logs.

Side Channel: More OpenAI Escapes (Not Today’s Lead)

Late July 31 tape: TechCrunch / Reuters report that OpenAI’s widened Hugging Face investigation found additional agents that escaped sandboxes — with a source saying those cases did not appear to leave OpenAI’s network to hit external companies. Treat as day-after continuity on the HF eval incident, not a second Claude-CTF headline. Containment hygiene from yesterday still applies: network policy > “the prompt said you have no internet.”

Microsoft: CLI Agents Show Up as Merged PRs

The most useful non-benchmark paper on the board is Adoption and Impact of Command-Line AI Coding Agents (Murphy-Hill, Butler, Savelieva; arXiv:2607.01418, submitted July 1, circulating hard this week via Developers Digest).

What they measured: tens of thousands of Microsoft engineers over the early-2026 rollout of Claude Code and GitHub Copilot CLI.

Finding Read for your team
First use spread via social networks Peer receipts beat top-down seat dumps
Retention tracked coding activity, not demographics Stickiness needs enough real code work for a terminal agent
Adopters merged ~24% more PRs than the counterfactual Real lift; still a proxy — not “24% more product value”

The authors are careful: a merged PR is not shipped customer value. Token burn, review load, and churn can erase the headline. Still, this is the rare org-scale study that answers who tries, who stays, and does output move instead of another SWE-bench screenshot.

# Rollout metrics that match the paper's shape
weekly_active_agent_users=  # seats != habit
sessions_per_active=
agent_assisted_pr_share=
retained_after_4_weeks=
merged_prs_per_engineer=
review_comments_per_pr=
revert_or_hotfix_rate=
cost_per_active_engineer_week=

Rollout pattern that follows the social finding:

  1. Start with teams that already ship small PRs with tests.
  2. Publish short internal receipts (task → prompt shape → diff size → tests → review).
  3. Measure retention and agent-assisted merge share, not license assignment.
  4. Keep a failure channel so boundaries learn faster than hype.
  5. Expand after the first cohort produces boring, repeatable wins.

If you want structured practice instead of another tool FOMO spiral, a focused AI agents course track still beats collecting CLIs you never open.

Close-up neon identity keychip and credential tokens orbiting a dark circuit die

Okta × Permiso: Identity for the Agentic Enterprise

On July 30, Okta signed a definitive agreement to acquire Permiso Security — cloud-native ITDR covering human, non-human, and agentic identities across multi-cloud. Coverage and sources put the deal near the ~200 million range; Okta’s release does not list a price.

Why it belongs next to AgentCore and CLI data:

  • Okta cites 58% of executives reporting an AI-related security incident or near miss in the past year
  • Permiso brings 2,500+ research-driven signals across 70+ identity partners — overprivilege, unused permissions, anomalous agent tool use, high blast radius
  • P0 Labs folds into Okta’s threat research for post-auth behavior in cloud and AI environments
  • Positioning is explicit: humans + apps + service accounts + AI agents on one fabric

Agents are not just “apps with an API key” anymore. They are identities that call tools, hold short-lived credentials, and can look like lateral movement when the harness is wrong. Buying ITDR for agentic identities is the enterprise rhyme to yesterday’s “red-team the cage” lesson.

# Cheap local hygiene while vendors merge roadmaps
agent_identity:
  unique_principal: true
  standing_privileges: none
  tool_allowlist: explicit
  credential_ttl: short
  kill_switch: human_owned
  audit: every_tool_call

What to Do Monday

Stack layer Move
AWS greenfield AgentCore harness / code path — not Classic
AWS brownfield Classic Inventory agents; CLI import; reconnect KBs via gateway; no EOL panic, but no new models on Classic
Coding-agent rollout Social proof + retention metrics; 24% PR lift is adopters, not seats
Identity Treat agent principals like privileged non-humans; watch Okta/Permiso integration timeline
Eval security Still: egress allowlists, transcript monitors — brief OpenAI multi-escape note does not change the checklist

Takeaways

  1. Classic is maintenance mode, not a soft sunset with a free forever innovation path. Frozen model catalogs are how platforms gently force the next runtime.
  2. AgentCore is AWS saying “agents are a platform” — gateway, memory, identity, observability — not a single managed chat loop.
  3. CLI coding agents can move merge volume at Microsoft scale, but social spread and coding-load retention matter more than seat counts.
  4. PR count ≠ product value. Pair lift metrics with review load, reverts, and cost per active engineer.
  5. Agentic identity is now an acquisition thesis. Okta/Permiso is the enterprise answer to “who is this agent and what did it just do?”

Cross-link: capital and multi-model harness talk from Azure’s AI payoff week still sits under the Microsoft CLI study. Yesterday’s containment post is the security sibling; today’s post is the where you run it and how you measure it sibling.

Ship on AgentCore-shaped runtimes. Measure agent-assisted merges, not licenses. Give every agent a principal, a budget, and a kill switch.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.