Architecture
This page is a technical deep dive into the AWS services Tessera uses and how each
feature uses them. It is service-focused; for the conceptual run-path walkthrough see
How it works, and for the infrastructure-as-code detail (Terraform
modules, deployment) see the repository's docs/architecture/ and docs/deployment/.
Service inventory
| Service | Role in Tessera |
|---|---|
| Amazon Bedrock AgentCore — Runtime | The single container that runs every agent, invoked per agent per run. |
| Amazon Bedrock AgentCore — Gateway | Exposes MCP tools to agents. |
| Amazon Bedrock AgentCore — Memory | Shared, namespace-isolated memory agents read/write within a run. |
| Amazon Bedrock — Guardrails | Contextual grounding checks on agent output (internal grounding). |
| Amazon Bedrock — Evaluations | LLM-as-judge scoring for the evaluation harness (bedrock_judge). |
| Amazon Bedrock AgentCore — Evaluations | Optional agent trajectory scoring in evaluation. |
| AWS Step Functions | One state machine per ecosystem (the agent DAG); plus catalog build and evaluation state machines. |
| AWS Lambda | Control-plane handlers, agent-invoker, catalog ingest/synthesis, and the evaluation Lambdas. |
| Amazon API Gateway (REST + WebSocket) | The JWT-authorized control plane and the live run stream. |
| Amazon S3 | Ecosystems, run artifacts, golden sets/scorecards, the static frontend, and access logs. |
| Amazon DynamoDB | WebSocket connections, the human-attention inbox, and catalog refresh state. |
| Amazon CloudFront + AWS WAF | Delivers the static frontend; WAF protects the distribution. |
| Amazon Cognito | User authentication (user pool, app clients, M2M client). |
| Amazon Verified Permissions | Per-ecosystem authorization with Cedar policies. |
| Amazon EventBridge | Fires online evaluation when a run completes; catalog schedules. |
| AWS Secrets Manager + AgentCore Identity | Zero-static-secret credential storage (e.g. the web-search key). |
| AWS KMS | Encryption keys for data at rest. |
| Amazon VPC | Network isolation for the Lambdas. |
Feature → services
Running an ecosystem
The WebSocket API starts the ecosystem's Step Functions state machine; each agent is invoked on the AgentCore Runtime, using Gateway tools and shared Memory. Grounding uses Bedrock Guardrails; artifacts land in S3.
Catalog ingest & synthesis
The catalog build state machine ingests sources, normalizes/synthesizes them via the AgentCore Runtime, tracks refresh state in DynamoDB, and promotes versions in S3. See Catalog.
Evaluation
The Eval State Machine runs the ecosystem headlessly over the golden set, scoring with deterministic scorers plus Bedrock/AgentCore Evaluations, and writes scorecards to S3. Online eval is triggered off completed runs by EventBridge. See Evaluation.
Quality & grounding
Auth & authorization
Cognito authenticates; API Gateway validates the JWT; Verified Permissions authorizes each ecosystem action against Cedar policies. See Share an ecosystem.
Frontend delivery
Data & state
| Data | Store |
|---|---|
| Ecosystems (YAML) | S3 |
| Run artifacts / reports | S3 |
| Golden sets + scorecards | S3 |
| WebSocket connections | DynamoDB |
| Human-attention inbox | DynamoDB |
| Catalog refresh state | DynamoDB |
| Agent working memory | AgentCore Memory |
| Static frontend | S3 (served via CloudFront) |
Security posture
- Network — Lambdas run in a VPC; the frontend is fronted by CloudFront with AWS WAF.
- Encryption — data at rest is encrypted with KMS; S3 buckets are private (CloudFront uses Origin Access Control).
- Authorization — every ecosystem action is authorized per-ecosystem with Amazon Verified Permissions (Cedar).
- Zero static secrets — credentials (e.g. the web-search key) live in Secrets Manager / the AgentCore Identity token vault, never in the repo or environment files.
This page covers the services. The Terraform modules that provision them, the
dependency graph, and deployment steps are in the repository's docs/architecture/ and
docs/deployment/.