// Production / Agentic Workflows & RAG
AI agents on AWS AgentCore.
AgentCore is the native runtime for agents that have to stay inside your AWS perimeter - compliance, data residency, the security review you've already passed. We build the agent layer in your account, with your engineers.
// What we see
Console works. Production breaks. Always in the same places.
01
Your supervisor adds 3 seconds per hop
Multi-agent collaboration looked clean in the diagram. In production, every supervisor → worker → return cycle adds 2-5 seconds - a 4-hop conversation that demoed at 8 seconds answers in 30.
02
Memory bleeds across users
Long-term and short-term memory are configured, but the scope is wrong somewhere - User A's context shows up in User B's responses. At compliance scale, that's an incident report.
03
Bedrock costs surprise you at scale
Cost per query was fine in dev. At 50 RPS the bill is 8x what you modeled - every supervisor hop is a separate billed Bedrock call.
// Case Study
Embedded analytics agent - multi-marketplace data, AWS Bedrock + Agent Core
A $50M+ ARR ecommerce data SaaS wanted their end-customers to ask precise, cross-marketplace questions in plain language - without an analyst in the loop. We built a dedicated-toolkit agent on AWS Bedrock + Agent Core: typed tools per question pattern, SQL kept only as a long-tail escape hatch. Robust where text-to-SQL falls over. 3 months from kickoff to production.
3 months
concept to production
5+
marketplaces unified
AWS
Bedrock + Agent Core, in-perimeter
// What we do
Three decisions that decide whether AgentCore holds up.
Most AgentCore deployments don't fail on the model layer. They fail on memory shape, supervisor topology, and observability that doesn't see what's actually slow.
Memory shape decides the project
Almost every production AgentCore problem traces back to a memory choice made in week one.
- Long-term, short-term, and session tiers mapped early
- Scope, retention, cost, and data-residency per tier
- Memory tier doubles as a compliance decision
Supervisor latency is a design constraint
Supervisor → worker hops cost 2-5 seconds each. Your latency budget shapes the architecture.
- Fewer hops by design
- Parallel workers where the conversation allows
- Tool calls inlined where AgentCore semantics permit
Observability inside your AWS stack
Native AgentCore + CloudWatch shows what fired. We layer what failed quietly on top.
- Trace replay and per-tool latency breakdowns
- Cost-per-conversation dashboards
- Feeds your existing CloudTrail and audit pipeline
// Method fit
AgentCore isn't the right runtime for every agent.
skip it if
You're not on AWS
AgentCore's value is the AWS perimeter - VPC, IAM, GuardDuty, CloudTrail, KMS. On GCP or Azure the plumbing connects to nothing you own.
Your data can leave AWS
If LLM calls can leave your AWS perimeter, direct OpenAI, Anthropic, or multi-cloud gateways open up - AgentCore's value is the audit boundary.
Your team owns the runtime
AgentCore is opinionated. An SRE team that wants to own retries, queue semantics, and tool dispatch gets more freedom on LangGraph + EKS.
LangGraph Agents
use it if
AgentCore fits when you're already on AWS, your compliance perimeter is the determining factor, you have approved-model access through Bedrock, and you want a runtime your security team has cleared.
// How we work
Architecture first. Iterate in a shared workspace. Hand off the dashboards.
Every AgentCore engagement starts with the architecture decisions that are expensive to undo - memory tier, supervisor topology, identity boundaries. From there we build in your AWS account, with your engineers in the loop the entire time.
01
Architecture review (week one)
We map the agent graph with your team - supervisors, workers, memory scope, tool boundaries, identity flow. Output: a written design your engineers approve before any code ships.
02
Build in your AWS account
All work happens in your account, your VPC, your IAM. Your engineers share the same Bedrock workspace, CloudWatch dashboards, and trace tooling - no vendor sandbox to migrate out of later.
03
Hand off cost and latency dashboards
Code, agent definitions in IaC (CDK or Terraform), the eval suite in your CI, and dashboards for your on-call. Slack support for 30 days after delivery.
// Expert insight
“AgentCore bills CPU-seconds and peak memory, not per invocation. Most of an agent's runtime is waiting on Bedrock - memory billed, CPU not. Smaller memory footprint matters more than fewer calls.”
Michał Świędrowski
Co-founder @ bards.ai
// Why bards.ai
Every change ships with a measured delta.
You could hire the team. It would take a year and they'd learn AgentCore on you - on your bill, in your audit trail. We've already learned it.
Production agent deployments at scale
Brand24's internal agent unifies 13 data sources, sub-5s p50. The patterns transfer: memory scoping, supervisor design, observability your on-call can debug mid-incident.
Eval-first methodology
Every agent change ships with a measured delta - latency, cost-per-conversation, task completion. Shared dashboard, no silent regressions to land a cosmetic win.
AgentCore learned before your engagement
Every person on your engagement has shipped agents inside an AWS perimeter - memory scoping, supervisor topology, and Bedrock cost mapping included.
// FAQ
Common questions about AgentCore
Bedrock Agents was the previous generation - a managed agent service tied to Bedrock LLMs with a fixed action-group model. AgentCore is the runtime layer underneath it, exposed: Memory, Identity, Gateway, Browser, Code Interpreter as separate primitives. AgentCore gives you more control and composability; Bedrock Agents is faster to demo but harder to bend. New projects with non-trivial agent logic should start on AgentCore.
LangGraph wins when you need full control of state, retries, and tool dispatch, or when AWS-specific compliance isn't a driver. AgentCore wins when you're AWS-first, need VPC/IAM isolation as a non-negotiable, and want a managed runtime that handles agent lifecycle, scaling, and durability without you owning the cluster. Both can coexist - LangGraph as inner-loop orchestration inside an AgentCore-hosted worker is a pattern we've shipped.
Three things: (1) Map the conversation tree to the pricing tree before any code lands - every supervisor hop is a billed call, and the model choice per hop matters. (2) Use cheaper models (Claude Haiku, Nova Lite) for supervision and routing, premium models (Sonnet, Nova Pro) only where the user-facing answer needs them. (3) Cost-per-conversation dashboards from day one, alerted on regression. The goal is to know the unit economics, not to discover them after the bill arrives.
// Related services
Adjacent problems we solve
- Learn more
Agentic Workflows & RAG
LangGraph Agents
Postgres checkpointing, recursion budgets, and eval gates that keep LangGraph agents alive under real traffic - on your infra, operated by your team.
- Learn more
Agentic Workflows & RAG
Enterprise RAG Pipeline
Your demo answered 30 questions; customers ask question 31. Retrieval, evals, and regression gates that catch it before they do.
- Learn more
Agentic Workflows & RAG
Agent Evals & Observability
Traces and evals that tell you whether retrieval missed, the model hallucinated, or a tool broke - and page someone when it matters.
// Let's ship it
An architecture your security team can approve.
Send the conversation shapes, your AWS perimeter, and the latency and cost bars - you'll get back an architecture and eval plan within a business day.
- Engagements from
- $40K
- Typical range
- $40K-$120K
- Duration
- 4-8 weeks
Fixed-fee proposal after the first scoping call. Scope drivers: number of agents, memory scope complexity, identity integration depth, multi-region scope.
Michał Świędrowski
Co-founder @ bards.ai