Agent loop / planned
Planned agents that turn signals into next actions.
Six agents, each with a narrow question, bounded evidence, and an explicit output. The first three ship at MVP; the rest follow. The goal is useful diagnosis—not an opaque autonomous system making silent production changes.
ScopeRead-only first
EvidenceTrace-linked
ActionHuman-approved
StatusPlanned capabilities
A deliberate analysis loop.
Each agent is organized by operational question. Their findings will link back to the evidence and remain reviewable before any downstream action to the Console, a notification, or an MCP response.
- 01
CostSpikeDetector
Where did spend spike?
Detect a rolling cost anomaly against the org’s baseline window, so a spend problem surfaces before the invoice does.
Planned output: ranked cost findings · MVP - 02
TokenOptimisationAdvisor
Which prompt is wasting tokens?
Find prompts with a wasteful prompt-to-completion ratio and suggest concrete compression.
Planned output: compression suggestions · MVP - 03
CachingOpportunityFinder
What should be cached instead of recomputed?
Cluster near-duplicate prompts that could be served from a cache instead of re-run at full cost.
Planned output: caching candidates · MVP - 04
PromptQualityScorer
Did the run get cheaper but worse?
Score output quality per prompt template so a regression from an edit, model swap, or temperature change doesn’t hide behind a lower bill.
Planned output: quality regression signal · post-MVP - 05
ModelDriftMonitor
Did the model change under you?
Track distributional change in model outputs over time, since providers update models and prompts age quietly.
Planned output: drift alert · post-MVP - 06
TrendForecaster
What will the bill look like next month?
Project cost and usage 7, 30, and 90 days forward from current trend, for budgeting rather than surprise.
Planned output: cost forecast · post-MVP