Advisor / Evidence-backed recommendations

See which AI workflows deserve review—and why.

Advisor turns connected cost and workflow signals into a focused queue: what changed, why it matters, the evidence behind it, and what to review next.

Start free → See what Advisor reviews →

Evidence attachedCalculations recomputedNo changes executed

A Recommendations for reviewEvidence, impact, and limits attached Illustrative
ContextCoverage: complete

Repeated context may not match recent work.

Compare standing repository instructions with the context recent sessions used.

Evidence currentImpact estimateInspect finding →
CacheCoverage: partial

Prompt churn may be reducing cache reuse.

Check write cost, reuse, and retention before changing the prompt.

Recurring workCoverage: complete

A recurring agent group may no longer match its work.

Compare cadence, parent task, and agent-family mix before deciding what to change.

What Advisor reviews

One bill. Many places to look.

Spend alone cannot tell you whether work is useful. Advisor compares how work runs with what it costs, then surfaces evidence-backed mismatches for review.

01

Context and configuration

See when instructions, tools, or configuration consume more context than the work appears to use.

02

Context composition

Compare the share of input used by repository instructions, tools, conversation, and other working context.

repotoolsconversationother
03

Prompt caching

Weigh cache writes and storage against reads that reused the prefix. A low hit rate alone is not a recommendation.

Write premium
Read discount
Storage
04

Repeated agent groups

Trace recurring and spawned sessions to their parent work. Compare cadence, work pattern, and agent-family mix where evidence supports it.

parent task scheduled spawned unattributed
05

Plan, API, and cloud allocation

Compare subscription coverage, metered API use, and cloud allocation by agent family—without assuming every source exposes the same detail.

Plan allowancecoverage observed
API + cloudusage observed
06

Provider impact

Connect provider pricing and model changes to the accounts and workflows they may affect, with dates and source health visible.

Publishedsource captured Effectiveimpact recalculated Reviewedrecommendation updated

All sample findings and values in this section are illustrative.

What every finding includes

See the evidence, not just the advice.

Each recommendation keeps its evidence window, source coverage, source health, confidence, assumptions, and caveats beside the conclusion. Missing evidence stays visible.

  • ✓

    Reviewed against the evidence. Customer-visible recommendations pass an evidence review gate.

  • ✓

    Bound to a clear window. Source dates, stale inputs, and missing coverage stay visible.

  • ✓

    Honest about uncertainty. Estimates stay labeled, and unattributed work stays unattributed.

Context composition Illustrative example

Standing repository context may exceed what recent work uses.

Review recurring instructions and tool definitions before changing configuration.

Bottom line
Candidate for review
Evidence window
Connected recent sessions
Coverage
Complete for this example
Confidence
Reviewed
Source health
Current
Caveat
Aggregate categories only
Repository instructionslargest
Tool definitionsreview
Conversationbaseline
Next stepInspect the evidence and decide whether to change configuration →
A governed recommendation loop

Analyze. Recompute. Review.

Advisor analyzes connected evidence and recomputes the finding before it enters the review queue. It never makes the change for you.

  1. 01
    Gather connected evidence

    Usage, cost, plans, context, configuration, provider changes, and source health.

  2. 02
    Analyze within guardrails

    Governed tools and bounded playbooks evaluate only the available evidence.

  3. 03
    Recompute the finding

    Calculations run again from fetched evidence; missing inputs remain explicit.

  4. 04
    Review the recommendation

    Inspect the evidence, assumptions, estimated impact, caveats, and suggested next step.

Advisor stops at the recommendation. It does not change plans or configuration, execute workflow changes, promise savings, or roll anything back. Next steps vary by finding.

Why workflow evidence matters

Agent spend is a systems problem.

The useful unit is the workflow: its context, tools, schedules, loops, and agent fan-out. A request-level bill alone misses how those parts interact.

Context engineering

Keep context useful.

Useful context has to be selected and maintained as sessions and workflows grow.

Explore Anthropic’s guidance ↗ Tool systems

Tool catalogs spend context.

Tool definitions can consume context before the agent begins the actual task.

Explore Anthropic’s guidance ↗ Agent fan-out

Parallelism changes the cost curve.

Fan-out adds work as well as capacity, so orchestration belongs in the review.

Explore Anthropic’s research ↗
Review your first finding

Know why a workflow is costly before you change it.

Connect the evidence. Let Advisor analyze and recompute. Review the recommendation on your terms.

Start free → Explore supported sources →

Recommendations are reviewed · impact is estimated · Advisor does not execute changes