Simulation & episodes

Episodes

When Eldros tests your agent, it opens a conversation and plays a scenario against it — through your normal protocol, exactly like a real user. One such simulated conversation is an episode. Every episode has an id, and the platform uses it to assemble the run’s results: transcript, judgments, and any traces that carry it.

episode (one simulated conversation / test run)
└── session (the conversation Eldros opened with your agent)
    └── turn / trace spans — stamped episode.id, traffic_type=simulation

This requires no code on your side. Eldros drives the conversation and captures the transcript itself, and it records which conversation belongs to which episode at the moment it connects. Your simulation results appear on the platform either way.

Joining your internal traces to an episode

If your agent runs the SDK, you can go one step further: tag the telemetry your process produces — turn spans, LLM and tool calls — with the episode, so the run’s results include your agent’s internal view. Eldros sends standard W3C traceparent + baggage headers on every connection it opens to your agent — an HTTP request, or the WebSocket handshake (the upgrade request is a plain HTTP GET). Continue them with one call, in any framework:

# HTTP: per request
with eldros_sdk.trace_context(request.headers):
    with eldros_sdk.turn(msg, session_id=conversation_id) as t:
        ...

# WebSocket: once after accept — one connection is one conversation
async def ws_handler(websocket):
    with eldros_sdk.trace_context(websocket.headers):
        while ...:
            with eldros_sdk.turn(msg, session_id=conversation_id) as t:
                ...

trace_context takes any header mapping — it is a plain context manager, so it works in every server framework, ASGI or not. Inside the block, every span gets:

episode.id — on every span, so the platform can fetch the run’s full telemetry with one filter.
traffic_type="simulation" — simulated traffic never pollutes your production views.
Session grouping for free — when no session id is present, spans group under the episode id automatically.

On production traffic these headers simply aren’t there, and trace_context is a harmless no-op passthrough — you can keep the same handler code for both.

When the id arrives another way

If your protocol carries the episode id somewhere else — a query parameter, a bootstrap message field — set it directly instead:

eldros_sdk.set_episode(episode_id)        # once, at connection start

# or scoped, when one process handles many episodes:
with eldros_sdk.episode(episode_id):
    ...

Set it before opening turns: identity is stamped when a span starts, so an episode set mid-turn misses the turn span itself.

Getting started

Capture

Integrations

Concepts

Simulation & episodes

Episodes

Joining your internal traces to an episode

When the id arrives another way

​Episodes

​Joining your internal traces to an episode

​When the id arrives another way

Episodes

Joining your internal traces to an episode

When the id arrives another way