Context

The fleet-wide context and memory picture: which agents have the most runs near the context limit, the largest share of context spent on tool definitions, or memory that isn’t helping. This is a fixed 30-day (and 7-day sparkline) ranking view with no time-range control of its own — open an agent’s own Context tab (Context & Memory) for the deep charts, prompt-cache panel, and memory-operations detail this page doesn’t carry.

What you see

Context & memory posture

Six earned-severity cards: Most runs near context limit, Largest tool definitions share, Instrumentation Coverage, Largest Context User, Memory Cost, and Memory Helpfulness. The first two name the measured share, for example “support-bot has the most runs near the context limit: 18% of its runs used more than 80% of the model’s context window.” A card is only colored (teal/amber/red) when its number is inherently good or bad — Largest Context User and Memory Cost just name the biggest number, so they stay uncolored. Cards without confident data move into a separate “Instrumentation gaps” list instead of showing a colored zero.

Fleet context growth and composition

A daily p50/p95 area chart of context growth across the fleet, followed by a composition trend chart.

Agent ranking

One row per agent: avg tokens, Tool definitions share, Runs near context limit, a 7-day sparkline, avg cost, and an efficiency score. Sortable by avg tokens or efficiency score.

Three metrics that look similar but aren’t: Runs near context limit is the share of runs that used more than 80% of the model’s context window; Tool definitions share is tool-definition tokens ÷ total context tokens; efficiency is utilization × (1 − tool definitions share). A cell shows a reasoned dash (hover for why) when a metric can’t be computed instead of a fabricated number — e.g. tool definitions share with no tool-definition data, or efficiency with no LLM utilization recorded.

Trace evidence

Seven categories of drill-down trace links backing the table’s numbers: highest context runs, the worst runs near the context limit (listed as Worst bloat examples), retrievals on failed runs, high-cost memory runs, traces where memory helped or may have hurt, and traces with unused tool tax.

Morse has no per-retrieval relevance score: retrievals on failed runs are shown as a place to look, not as evidence that a retrieval caused the failure.

How to use it

Sort the ranking table by avg tokens or efficiency score.
Click any ranking row to open that agent’s Context tab (Context & Memory), carrying a ?from=context-intelligence breadcrumb back to this page.
Hover a column header’s ⓘ icon for its exact definition.
Expand a trace-evidence category to see the runs backing it, and click one to open its trace.

Before you have data

If no agent has context data for the past 30 days, the ranking table shows: “No context data for the past 30 days. Agents appear here once they have traces with LLM spans. If you have recent traces, make sure you’re logged in as the account that owns them.”

Agents context tab — this page’s per-agent drill-down.
Agents — the fleet view for health and cost instead of context.