Ecosystems

An ecosystem is a multi-agent workflow declared in YAML. It's the central artifact you author in Tessera: it names the agents, what each one does, how they depend on each other, what inputs the operator supplies, and what the run produces.

When you save an ecosystem, Tessera compiles it into a single Step Functions state machine from the agent dependency graph.

Anatomy of an ecosystem

ecosystem:
  name: Property Buyer Report
  description: Takes a property address, researches it, produces a due-diligence report.
  version: "1.0"

strategy: orchestrated      # orchestrated | choreography
report: true                # auto-injects a report_agent
quality:
  challenger: true          # auto-injects a challenger_agent
  grounding: true           # internal + external grounding checks

inputs:
  - id: address
    type: text
    label: Property Address
    required: true

entry_point: research_agent

memory:
  strategies: [semantic]    # add 'episodic' for chronological recall
  shared_namespaces:
    - name: findings
      write: [all]
      read: [all]

agents:
  - id: research_agent
    name: Property Research
    swarm: research
    tools: [web_search, memory_write, current_time]
    output_kind: analyst
    prompt: |
      Research the property from public sources and extract structured facts...

  - id: report_synth
    name: Report Synthesizer
    swarm: synthesis
    depends_on: [research_agent]
    tools: [memory_read, code_interpreter, generateUI]
    output_kind: synthesizer
    prompt: |
      Reconcile the upstream findings into a single buyer report...

swarms:
  research:  { agents: [research_agent], strategy: graph }
  synthesis: { agents: [report_synth], strategy: graph }
parallel_swarms: [research, synthesis]

triggers:
  - type: manual

See the Ecosystem YAML reference for every field.

Top-level settings

Field	Purpose
ecosystem.name / description / version	Identity shown in the UI.
strategy	orchestrated (DAG) or choreography.
report	When true, the compiler adds a report_agent that assembles the final deliverable.
quality.challenger	When true, adds a challenger_agent that stress-tests the output.
quality.grounding	Enables internal (Guardrail) + external (web search) grounding checks.
inputs	Operator-facing fields (text, select, file) shown in the launch UI.
entry_point	The agent the run starts on.
memory	Shared namespaces agents read/write to pass context.

note

report_agent and challenger_agent are auto-injected by the compiler when report: true and quality.challenger: true. Do not define them in agents[] yourself.

Strategies

• orchestrated — agents run as a directed acyclic graph. Leaf agents (no depends_on) run first, in parallel; synthesizers wait for their dependencies. This is the common case.
• choreography — looser, event-driven coordination (used by a subset of ecosystems).

Where ecosystems live

Bundled ecosystems ship under app/content/ecosystems/. make seed publishes and compiles them so they appear in the UI. You can also create and edit ecosystems directly in the UI (including with the AI-assisted builder) — see Build an ecosystem.