Bedrock Agents Classic → AgentCore Migration Cost: The Platform Bill Barely Moves — Your Labor Is the Number

On July 30, 2026, AWS stopped opening Amazon Bedrock Agents to new customers. The service is now Amazon Bedrock Agents Classic, the model catalog is frozen, and a chunk of the coverage that week called it "killed." The primary source says something different — maintenance mode, no end-of-life date, no migration deadline [1]. For an AI agency, that gap is where the money is: what does a Classic → AgentCore migration actually cost, and what should you quote a client?

Most coverage in this cycle decodes AgentCore's price list — roughly a dozen independent meters — and stops there. This page does the inverse. We modeled the migration with AWS's published AgentCore pricing [3] and the one first-hand migration report we could find [10], and the result inverts the scare story: the platform delta is noise. Rework labor is the number. Run your own stack through the AgentCore Migration Estimator on the calculator homepage.

The Short Version (TL;DR)

• Classic itself is free — you paid only for model inference [1]. AgentCore bills ~12 components: runtime, gateway, memory, web search, policy, observability [3].
• First-hand report: a simple single-agent migration took ~4 hours; the monthly bill moved ~$4.87 → ~$5.02 (+$0.15) because tokens dominate and AgentCore's harness is more token-efficient [10][1].
• The real cost is labor: hours (simple) → 1–3 weeks (custom orchestration) → 3–6 weeks (multi-agent) [9].
• No EOL, no deadline — migration is a strategy choice, not an emergency [1].

Step 0: Are You Even Forced to Migrate?

Before you quote a migration, quote the decision. AWS's own doc is explicit that nothing forces an allowlisted account to move today [1]. What the freeze actually blocks is narrower than the coverage suggests:

• CreateAgent and InvokeInlineAgent return HTTP 403 AccessDeniedException for accounts without Bedrock Agents usage in the prior 12 months — the exact error: "Bedrock Agents is in Maintenance Mode. New agent creation is not available for accounts without prior service usage." There is no exception process [1][6][12].
• Everything else keeps running: update/get/list/delete APIs, InvokeAgent, action group/KB/alias APIs remain available, and allowlisted workloads run in maintenance mode with no end-of-life date [1][6].
• The model catalog is frozen as of July 30, 2026 — new models released after that date are available through AgentCore only [1][8].

The 6–12 month model-upgrade trigger

Here is the practical forcing function: any client model roadmap that needs a model released after July 30, 2026 requires the move. A client on a current model with no upgrade planned can stay on Classic. The moment a newer model enters the roadmap — typically within a 6–12 month upgrade cycle — AgentCore becomes compulsory for that workload [1][9]. Put that trigger in front of clients: the freeze is about the catalog, not the calendar.

How AgentCore Pricing Works: The Components, Decoded

AgentCore replaces Classic's per-invocation model with consumption-based pricing across ~12 independent components [3][4]. The ones that show up on an agency's bill:

Runtime — $0.0895/vCPU-hour + $0.00945/GB-hour, active use only

You pay for compute while an agent session is actively running; I/O wait is free [3]. This is the steepest learning curve for agencies used to per-invocation thinking — the bill tracks active vCPU-hours and memory, not request counts.

Gateway — $0.005 per 1K invocations

MCP tool exposure runs through the Gateway; every tool invocation is a metered call [3]. A chatty agent can generate dozens of gateway invocations per session.

Memory — $0.25/1K short-term events; $0.75/1K records/month long-term

Short-term memory events meter per 1K; long-term records meter per 1K stored per month [3]. Memory-heavy agents — conversation history, state, retrieval — gain a real line item they never had under Classic's free platform.

Web Search — $7.00/1K queries; Policy auth — $0.000025/request

Web search is a flat per-query meter; policy authorization requests are effectively noise at agency scale but are billed [3]. Observability moves to a new CloudWatch namespace [11].

The meter that dominates both stacks: model inference tokens

Both stacks share the dominant cost: model inference tokens. Classic billed per invocation; under AgentCore the token bill is largely unchanged — and AgentCore's harness is more token-efficient than Classic's internal prompts, so AWS's own claim is "comparable or lower" inference cost [1][10]. That claim is AWS's, not a guarantee — but it explains why the worked example shows a platform delta measured in cents.

What to collect before you estimate

For an agency, the practical takeaway is that the AgentCore bill is a function of four inputs you already control: active runtime hours, tool-call volume, memory events, and web-search volume. Collect those four numbers from a client's existing agents — or estimate them from session patterns — and the platform bill is arithmetic. The AgentCore Migration Estimator does that math from exactly these inputs, so the quote and the client's eventual AWS bill are built from the same assumptions.

Cost Comparison: Classic vs AgentCore, Line by Line

The worked example deserves its framing: it is one first-hand report from a DEV.to AWS Builder, not an AWS guarantee or a benchmark [10]. Directionally it matches AWS's own claim that AgentCore's harness is more token-efficient than Classic's internal prompts [1]. For a quote, treat the platform delta as roughly neutral and spend your estimating effort on labor.

Read the table as a meter inventory, not a price list to memorize. The only lines an agency should model per client are runtime hours (agents × sessions × active runtime), gateway invocations (tool calls), memory events, and web-search volume — plus the shared inference tokens, which dominate. Everything else on the AgentCore price sheet is rounding error at agency scale, and modeling it separately is how migration scare quotes get inflated.

Where the Real Cost Hides: Architecture Rework

The price list is the visible part. The expensive part is what has to be rebuilt or re-plumbed when a Classic agent becomes an AgentCore agent [1][4][11]:

• Action groups → Gateway/MCP or code-level @tools. Classic action-group wiring doesn't port as-is [1][4].
• Memory → AgentCore Memory, and IAM roles need reconfiguration. Managed Classic memory becomes the billed AgentCore Memory component, and the permission model changes [1][4].
• Breaking API change: client code must switch InvokeAgent → InvokeAgentRuntime — a code change in every downstream caller, not a config toggle [1][11].
• What does NOT replicate: stage-specific prompt overrides are not directly replicated in the AgentCore harness — agencies relying on per-stage prompt control need custom code. Multi-agent supervisor routing is also described as "not straightforward today" [4][11].
• Observability moves to a new CloudWatch namespace, so dashboards, alarms, and billing attribution built on the old namespace must be rebuilt [11].

One flagged item: a vendor blog (ServerGurus) warns that "a poorly architected multi-agent setup can 3x your AgentCore bill compared to a single-agent Bedrock Agents deployment" [7]. That is a single-source vendor warning, not a verified fact — a reason to architect carefully, not a pricing input.

Plan the migration as a sequence, not a cutover: audit the account allowlist, inventory action groups and prompt overrides, map memory and IAM, switch the API call in client code, re-point observability, and only then move traffic. Most of the cost is not the move itself — it is discovering mid-migration that a stage-specific prompt override or a supervisor-routing assumption does not port. That discovery is exactly what the labor ranges below are paying for.

Migration Labor: The Number That Actually Moves Your Quote

Here is the labor picture from the community migration reports [9], the first-hand data point [10], and AWS's own estimate [1][8]:

• Simple agents (model + action groups + knowledge base): hours — AWS says CLI import is fast [1]; the community table says hours–1 day [9]; the first-hand report landed at ~4 hours [10].
• Custom orchestration: 1–3 weeks [9].
• Multi-agent complex graphs: 3–6 weeks — flagged: the "~6 weeks" community reports are single-source [9], and AWS's secondary estimate is 2–4 weeks for most workloads [8]. Present the range, don't promise the point.

At a blended agency rate of $150–$350/hour (the band this site's cost of hiring an AI agency guide documents), the labor swing is the difference between a ~$600–$2,800 engagement (simple, 4–8 hours) and a ~$45,000–$126,000 engagement (multi-agent, 3–6 weeks × 40 hours). That is the number that moves a quote — not the $0.15/month platform delta.

What agencies should quote

A defensible migration quote has four components: (1) a freeze audit of the client's AWS accounts, (2) a migration scope — which agents are simple, custom orchestration, or multi-agent graphs, (3) a labor estimate from the table above, and (4) an ongoing AgentCore bill estimate from the estimator. Package the audit as paid discovery — the 403 trap below is why it must happen before you quote.

The Fresh-Account 403 Trap (agency-specific)

The allowlist is automatic and per-account, based on any Bedrock Agents activity in the prior 12 months — and AWS says, verbatim, "No. There is no exception process." [1]. For agencies that deliver on per-client AWS accounts — the most common delivery pattern in this industry — that is a landmine [11][12]:

• Mixed allowlist status across accounts. A production account with usage history is allowlisted; a fresh staging account, sandbox, or brand-new client account is not. The same code that runs in prod 403s in staging [11][12].
• IaC modules fail silently on new accounts. CloudFormation/Terraform provisioning AWS::Bedrock::Agent fails on unallowlisted accounts — often mid-stack, after setup is already billed [11][12].
• No appeal. With no exception process, a net-new client account is a hard dead end for Classic agent creation. Any net-new Bedrock agent work on a fresh account must go to AgentCore from day one — that changes how you scope the quote, not whether you can deliver [1].

Audit allowlist status before quoting any net-new project: check whether the account has Bedrock Agents usage in the prior 12 months; if it doesn't, price the work on AgentCore rather than Classic. It is the same audit discipline the AI agent migration service playbook applies to any client stack move.

Frequently asked questions

Sources & Methodology

Primary source is AWS's own maintenance-mode documentation [1] and the AgentCore pricing page [3]. The $4.87 → $5.02 worked example is a single first-hand migration report [10], labeled as a case study, not an AWS guarantee. Labor ranges come from a community migration guide [9] and are presented as ranges; AWS's secondary 2–4 week estimate is separately attributed [8]. The "3x bill" warning is a single vendor blog [7], flagged as unconfirmed. Full source list:

• [1] AWS — "Amazon Bedrock Agents Classic maintenance mode": docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.html
• [3] AWS — Amazon Bedrock AgentCore pricing (runtime $0.0895/vCPU-hr + $0.00945/GB-hr; gateway $0.005/1K; memory $0.25/1K events, $0.75/1K records/mo; web search $7.00/1K; policy $0.000025/request): aws.amazon.com/bedrock/agentcore/pricing
• [4] RPABotsWorld — "Bedrock Agents Classic Sunset Migration Guide": rpabotsworld.com
• [6] aiagentstore.ai — "Daily AI Agent News — July 31, 2026": aiagentstore.ai/ai-agent-news/daily/2026-07-31
• [7] ServerGurus — "AWS Just Killed Bedrock Agents — What AgentCore Means" (single-source "3x bill" warning — unconfirmed): servergurus.com/blog/aws-bedrock-agentcore-2026
• [8] Enterprise DNA — "AWS Retires Bedrock Agents — AgentCore Is the New Path" (2–4 week secondary estimate): enterprisedna.co
• [9] byteiota — "AWS AgentCore Is Here — What to Do Before Classic Locks You Out" (labor table; single-source "~6 weeks" reports): byteiota.com
• [10] DEV.to (AWS Builder) — "My Robot Accountant Got an Eviction Notice…" (first-hand migration; $4.87 → $5.02; ~4 hours): dev.to/aws-builders
• [11] eCorpIT — "Bedrock Agents Classic — 30 July 2026 migration map" (InvokeAgent → InvokeAgentRuntime; CloudWatch namespace; IaC failures): ecorpit.com
• [12] CloudKaramchari — "Bedrock Agents Classic Enters Maintenance Mode" (403 behavior, allowlist, no exception process): cloudkaramchari.com

Accuracy note: AWS mechanics — the July 30, 2026 cutoff, HTTP 403 on CreateAgent/InvokeInlineAgent, the per-account 12-month automatic allowlist, "no exception process," the frozen catalog, no EOL date, Classic's free platform, and AgentCore component rates — come from AWS's maintenance-mode doc and AgentCore pricing page, verified 2026-08-16 and 2026-08-29. The $4.87 → $5.02 delta and ~4-hour migration are a single first-hand report, labeled as a case study; AWS's "comparable or lower inference cost" is AWS's own claim. The "3x bill" and "~6 weeks" items are single-source and flagged inline. Labor ranges are planning ranges, not AWS commitments. AgentCore model choice can be any provider, so token pricing may differ from Classic-era Bedrock rates — model your actual model.