Traces vs execution evidence for AI agents

When an AI agent run goes wrong a week later, the question is rarely "was the service up?" It is "what exactly did that run see and do, and can I look at it again now that the world has changed?" Logs and traces answer the first kind of question well. They were not built to answer the second.

A trace is a view of a run. A Run Capsule is an artifact you keep.

What logs and traces are good at

Observability tools such as Langfuse and LangSmith, or any OpenTelemetry backend, show you what is happening across many runs: latency, cost, error rates, a timeline of spans. That is the right tool for watching a system in production. NovaFabric does not replace it, and it can sit next to it. See the comparison and the NovaFabric vs Langfuse tutorial.

What a trace backend usually is, though, is a record in a store somebody else operates, with its own retention window and its own schema. When the window passes, the vendor changes, or the model behind the API is updated, the run is gone or no longer interpretable.

What a Run Capsule keeps instead

A Run Capsule is a plain directory on your machine. It holds the model calls (the request messages and the response text), the tool calls, the execution spans, the inputs and outputs, an environment snapshot, a replay policy and lineage edges, plus a manifest that records a SHA-256 digest of each evidence file. You can copy it, archive it, attach it to a ticket, or hand it to someone else. It needs no server and no account.

Capture wraps your command without code changes, and runs a built-in secret scanner (14 key and token rules) over the capsule's free-text streams. The scanner matches known patterns; it does not find every possible secret, so review a capsule before sharing it.

Replay is a different job from tracing

Because the capsule keeps the recorded model responses, a run can be replayed from it. There are five replay modes; intervention is experimental. In mocked replay, NovaFabric serves the recorded replies to synchronous OpenAI and Anthropic chat calls, so no live model call is made. Tools still run live. This is not deterministic replay of an arbitrary agent, and it is not an offline re-run of one. The replay page states exactly what each mode reuses and what it does not.

What sealing adds, and what it does not

A capsule can be sealed with your own key (opt-in, through a novaseal.yaml file) and verified later without a server. A valid seal establishes one thing: the recorded capsule is unchanged since the holder of that key signed it. Who the signer is depends on a trust anchor, which is an experimental feature today. A seal does not establish that the record is complete, because capture sees what it is wired to see. It does not establish that the record is true, because a key holder can sign a false capsule.

That is a narrower claim than "proof of what the agent did", and it is the one the software can honestly support. If you need the longer walk-through, the tutorial Prove what an AI agent did, months later, offline runs it end to end.

Which one do you need?

• Watching production behaviour across many runs: use your observability stack.
• Keeping one important run so you can reopen, replay, compare and verify it later: a Run Capsule.
• Both: they are complementary.

The capsule format is documented but pre-1.0, so expect additive changes until the freeze. To try it, start with Getting Started.