Native Observability

Manowar observability is native runtime data. It does not depend on LangSmith or an external tracing SaaS to know what happened during a run.

What Gets Recorded

Data	Where	Notes
Token usage	`telemetry.ts`	Extracts prompt, completion, reasoning, and total tokens from multiple provider shapes.
Usage records	`UsageTracker`	Records agent wallet, run id, model, action, source, and timing.
Run metadata	`upsertRunMetadata()`	Stores status, wallet, model, timing, error, and arbitrary metadata.
Feedback	`recordFeedback()` and helpers	Supports quality score, error, insight, and decision feedback.
Learnings	`recordLearning()` and `getRelevantLearnings()`	Lets workflow planning review prior outcomes.
Workflow checkpoints	`workflow/checkpoint.ts`	Records observations, decisions, insights, outputs, and errors.
MAL checkpoints	`harness/checkpoint.ts`	ValKey resume snapshots for typed MAL plans.

Usage Extraction

extractTokens() accepts provider output shapes that expose token fields under names such as:

prompt_tokens
completion_tokens
reasoning_tokens
total_tokens
inputTokens
outputTokens
reasoningTokens
totalTokens

resolveAuthoritativeTokens() prefers provider-reported usage when present and falls back to tracked callback totals. This matters because streaming and tool loops can expose usage at different points depending on the model adapter.

Run Lifecycle

Workflow runs use workflow/run-tracker.ts for lifecycle state:

Status	Meaning
`pending`	Created but not started.
`running`	Execution has started.
`success`	Completed with a result.
`error`	Failed with an error.
`cancelled`	Stopped by caller or runtime control.

Cron metadata records total executions, successes, failures, and last-run details per trigger.

Checkpoints Are Not All The Same

Checkpoint type	Storage	Purpose
Agent filesystem checkpoint	Local JSON files	Local checkpoint saver in `harness/checkpoint.ts`.
MAL checkpoint	ValKey `mal:run:{agentWallet}:{runId}`	Resume state for a typed plan.
Workflow annotation checkpoint	Memory/telemetry	Human-readable observations, decisions, insights, and errors.

Comparison

System	Observability model	Manowar model
OpenAI Agents SDK	Built-in tracing around SDK runs.	Native usage, feedback, and run metadata tied to Compose agents.
LangGraph plus LangSmith	Graph state traces and hosted trace UI when LangSmith is used.	Runtime-owned records and checkpoints without a graph dependency.
AutoGen	Event-driven agent logs and runtime events.	Wallet-scoped run records plus Compose metering evidence.

Overview

Memory

Harness

Tools

What Gets Recorded

Usage Extraction

Run Lifecycle

Checkpoints Are Not All The Same

Comparison

​What Gets Recorded

​Usage Extraction

​Run Lifecycle

​Checkpoints Are Not All The Same

​Comparison

​Related

What Gets Recorded

Usage Extraction

Run Lifecycle

Checkpoints Are Not All The Same

Comparison

Related