Introduction
Multi-agent AI workflows on AWS, declared in YAML — compiled to AWS Step Functions, run on Amazon Bedrock AgentCore, and streamed as live generative UI.
Tessera lets you build and run multi-agent workflows — called ecosystems — through configuration rather than code. You describe what agents should do and how their work fits together in YAML; Tessera compiles that into a runnable pipeline, executes it on AWS, and streams a generative UI back to a web frontend.
Ecosystems are domain-agnostic — the same engine powers a property due-diligence report, a commodity-trading desk, an enterprise risk board, an AI-governance assessment, a security incident war-room, and even a trip planner. What ties them together is how they run: agents grounded in a curated catalog, coordinated as a dependency graph, checked by built-in quality gates, and optionally measured by an evaluation harness. That machinery makes Tessera especially strong for regulated, high-stakes work where outputs must hold up — while staying general enough for anything you can express as collaborating agents.
Why Tessera?
- Configuration over code — an ecosystem is a YAML file. Add an agent, change a dependency, or swap an output format by editing config, not writing services.
- Multi-agent orchestration — agents run as a dependency graph across parallel swarms, each with a focused role, its own tools, and shared memory.
- Grounded, not guessing — domains ingest curated regulatory and reference sources into a versioned catalog that agents cite, backed by internal and external grounding checks.
- Defensible by design — built-in quality gates (output/format/citation checks, a challenger agent, and a quorum deliberation model), plus an opt-in evaluation harness for regression and model-risk sign-off, make outputs reviewable.
- Runs on AWS — Amazon Bedrock AgentCore for the agent runtime, Step Functions for orchestration, and a static web UI behind CloudFront.
Tessera vs. code-first frameworks
Code-first agent frameworks give you primitives and leave the hard parts — grounding, quality, UI, deployment — to you. Tessera is opinionated about exactly those parts.
| Tessera | Code-first frameworks (LangGraph · CrewAI · AutoGen) | |
|---|---|---|
| Define a workflow | YAML config | Write & maintain Python |
| Grounding | Versioned catalog + internal/external grounding checks | Build it yourself |
| Quality | Built-in challenger, quorum deliberation, format/citation/grounding gates | Build it yourself |
| UI | Live generative UI streamed to a web app | None / build your own |
| Deployment | Terraform to your AWS account (dev/staging/prod) | Bring your own infra |
| Secrets | Zero static secrets (AgentCore vault / Secrets Manager) | Varies |
| Evaluation | Opt-in eval harness (regression + model-risk sign-off) | Build it yourself |
Key concepts
| Term | Meaning |
|---|---|
| Ecosystem | A multi-agent workflow declared in YAML: agents, inputs, dependencies, and outputs. Compiled into one Step Functions state machine. |
| Agent | A single role in an ecosystem — a prompt, a set of tools, an output_kind, and its place in the dependency graph. |
| Swarm | A named group of agents that can run in parallel. |
| Domain | A knowledge scope (e.g. a regulatory area) that ingests curated sources into the catalog. |
| Catalog | The versioned, grounded knowledge store agents read from and cite. |
| Quorum | A deliberation model where a panel of scrutineer personas reach a verdict. |
| Evaluation | Offline capability eval + regression deploy-gate + online monitoring for an ecosystem. |
| Mission | The operator's request for a given run, supplied through the ecosystem's inputs. |
| Report | The final, structured deliverable produced by a run. |
How a run works
- An operator opens the web UI, picks an ecosystem, fills in its inputs, and launches.
- The WebSocket API starts the ecosystem's Step Functions state machine (compiled from the agent DAG when the ecosystem was saved).
- Each agent is invoked on the Bedrock AgentCore Runtime, using its tools and the shared memory. Agents emit generative-UI events that stream live to the browser.
- Quality gates run — format/citation checks, grounding, and (optionally) a challenger.
- A report agent assembles the final report, which the operator reviews and can give inline feedback on.
Quick links
- Installation · Quick Start
- How it works · Architecture · Ecosystems
- Build an ecosystem · Evaluate an ecosystem
- Ecosystem YAML reference
- GitHub repository
This site focuses on using Tessera. For the infrastructure and Terraform detail
(modules, state backend, CI/CD), see the repository's top-level docs/ directory.