Context

This agent’s full context and memory picture, as plain questions: what its context budget is spent on, how it’s growing, whether prompt caching is saving money, and whether its memory store is actually helping. This is the deep-dive companion to the fleet-wide Context ranking — that page tells you which agent to look at; this tab is where you look.

What you see

The tab sits under the agent’s tab row (it is the Context tab of the agent page). A 7d / 30d / 90d window switcher scopes the whole tab, carried in the URL and shared with the other agent tabs so switching doesn’t reset your range.

The tab is organized as plain questions, each with a one-line answer of what the section shows: “Three questions about what this agent sends to the model, answered from its recent runs”, then priced opportunities kept apart from them.

What is filling this agent’s context?

“Where the tokens go on a typical run: prompt, history, tool definitions, retrieved data.”

Context spend

Where this agent’s input tokens go — system prompt, history, tool definitions, retrieved data — with token counts per bucket, shown once source data exists.

Tool usage

Tools this agent has called, with defined-vs-called counts, call rate, and estimated monthly waste — when your SDK reports which tools were defined in the prompt, not just which were called. Stacks that only report calls lose those extra columns instead of showing fabricated numbers.

A detected-stack panel follows when the SDK reported one; it collapses when empty.

Is it getting close to the context limit?

“How big each run’s context is over time, and when it jumps well above this agent’s usual size.”

Context Growth & Stability

A token-size-per-run area chart with a 4× baseline line; runs that spike past it are marked with red dots you can click to inspect the offending trace. Above a 60-day window the chart switches to day-bucketed (max-per-day) points and drops the per-run drift bar strip, since a run-over-run delta stops meaning much once a point represents multiple days.

Is memory helping?

“What this agent reads from and writes to memory, and how runs with memory compare to runs without.”

Memory operations

Ops/day, a memory-operations timeline, context churn, and a cost/outcome comparison of runs with and without memory — but only for agents that have actually emitted a memory span. An agent with zero memory ops sees a quiet “This agent does not use external memory” card instead, not a section full of zeros.

The memory-on/off comparison itself needs 50+ traces per arm before it gives a verdict; below that it shows the trace counts and an insufficiency message rather than a comparison.

Where could it spend less?

“What changing it could cost less, as estimates. These are opportunities, not money already lost.” Priced suggestions live here, apart from the measurements above.

Recommended actions

Deterministic, evidence-linked suggestions for this agent. Accepting or dismissing one moves it to an “Accepted fixes” ledger; both actions are disabled (with a tooltip) in demo mode.

Prompt cache

Hit rate, estimated savings, write cost, and an invalidation timeline. Cache economics differ sharply by provider — Anthropic reads cost 10% (less on a few newer models) / writes cost 125%, OpenAI reads cost 50% / writes are free — and Morse applies the correct multiplier per model automatically, using the price catalog’s own cache rates when it has them, for both OpenAI and Anthropic spans, once cache tokens are present.

Agents on the Claude CLI subprocess adapter can’t expose cache control and show the cache empty state permanently — that’s an SDK-routing limit, not a pricing gap.

The hit rate is cache-read tokens divided by total input tokens (input already includes reads and writes), so 600 cached tokens out of 1,000 shows 60%.

Missed caching opportunity

For Anthropic agents that are not caching, Morse looks for a leading run of context (tool definitions, system prompt) that is identical across at least 60% of at least 20 calls and is long enough to be cacheable for that model. When fewer than 5% of those calls read from cache, the agent gets an “Enable prompt caching” recommendation and the same finding appears in the Cost Intelligence opportunities list. It states what was measured — for example “the first 9,000 tokens were identical across 240 of 240 calls; 0 were served from cache” — and a list-price savings estimate for the better of the 5-minute and 1-hour cache, named. It does not say why caching is off. OpenAI and Gemini agents are not checked yet. Morse reads content hashes and token counts only, never the prompt text.

How to use it

Switch the 7d / 30d / 90d window — it stays in sync with the other agent tabs.
Click a spike dot on the growth chart to inspect the offending trace.
Toggle the growth chart between hourly and daily granularity.
Accept or dismiss a recommended action to work through the fix list.
Click a sample-trace link in any evidence section to open that trace.

Before you have data

Sections degrade individually rather than the whole tab going blank: no source data shows an empty state on the Context spend card; no memory spans replace the whole memory block with the explanatory card above; no tool-definition data drops the affected table columns.

Agent detail — this agent’s Summary, Changes, Runs and Setup health tabs.
Context — the fleet-wide ranking this tab is the drill-down for.