Platform · Layer 03
Multi-model runtime, human in the loop.
Bedrock AgentCore and LangGraph run the agents, the routing picks the model per task, and a person stays on every action you flag.
07 components in this layer
- In production× 6
- Reference architecture× 1
Amazon Bedrock AgentCore
Managed runtime, identity and memory in your AWS account.
LangGraph orchestration
Every step, retry and hand-off is a node you can test.
Multi-model routing
Claude, GPT, Gemini, Nova, Llama and open-weights, by eval and cost.
Multi-agent orchestration
A supervisor delegates with typed hand-offs and one trail.
Human in the loop
Approval gates on anything you mark. Full context, one button.
Own voice stack
Parakeet, Qwen3 32B and Kokoro, with no per-minute vendor.
On-prem voice
The same stack on your GPUs, for calls that cannot leave.
In production: A real client, or DaCodes itself labeled as internal use, with a KPI measured in the Proof Ledger.
Runtime
What happens between a message and an action.
Six steps, whatever the channel. Each one is a node in the graph.
Receive
The channel adapter turns any channel into one event.
Look up
Tools over MCP read your systems with scoped credentials.
Reason
The router picks the model, the graph runs it with your rules.
Check
Guardrails, PII masking and thresholds, before any write.
Act or ask
It writes through scoped tools, or opens an approval with the evidence.
Log
Every step lands in the audit trail and the eval set.
00:00
Form submitted from /agents/collections-renewals
00:04
Enriched: 1,200 employees · insurance · México
00:09
Classified: qualified, 91% confidence
00:31
WhatsApp sent · reply received 02:10
03:48
Meeting booked with the account owner
Under five minutes
Time to first contact 0:31 · meeting in 3:48 · routed round robin
Own voice stack
Voice at a fraction of the market price.
Most voice agents die on the cost per minute. Ours runs on open models inside your cloud.
Good morning, this is the dispatch assistant for Route 18. Are you on schedule for the 14:00 delivery?
Running late, about twenty five minutes. Traffic on the highway.
Understood. I am moving the delivery window to 14:25 and notifying the customer. Anything else to report?
No, all good.
ETA updated in TMS
14:00 → 14:25 · customer notified on WhatsApp · dispatcher alerted
Own voice stack · about $0.035 per 15-min call, internal measurement, Aug 2026
- Parakeetspeech to text
- Qwen3 32Breasoning
- Kokorotext to speech
Per 15-minute call
- ElevenLabs
- $2.56 to $5.17
- Own stack
- $0.035
Per year, same volume, at group scale
- Per-minute vendor
- ~$5M
- Own stack
- ~$60K
Source: DaCodes internal measurement, Aug 2026. Comparison at list prices; volumes modeled for a 130,000-employee group in México.
Questions
Architects ask.
Talk to an architect
45 minutes with the people who run the platform. No slides.
We reply in under one business day.