Knowing what your AI agent did, and what it cost you
Field notes on cost observability for AI agents: what to store, what to alert on, and why raw tokens beat computed costs.
currently senior backend engineer at glitch ai
remote · utc+1
Eight-plus years of production Python. I design the systems behind the product, and lately the autonomous agents inside them: cost telemetry, audit trails, and self-review gates that make an agent accountable. Friends call me Caspian.
production systems
My work lives inside companies, in private repos and production infrastructure. So instead of a wall of repo links, each system is written up: the problem, the design, and how it held up.
An autonomous engineering agent in daily production use. It triages Sentry errors, answers engineering questions in Slack, reviews PRs, and opens draft fixes that it reworks when reviewers push back.
PythonSlackSentryGitHubRedisTelemetry that stores raw tokens so provider price changes never corrupt history, anomaly alerts with cooldowns, and investigation hints that point at the likely cause of a spend spike.
PythonPostgreSQLRedis160+ SEO-optimized topics per client domain across an enterprise client base, run on Celery/Redis pipelines with retries and failure recovery, on top of zero-downtime Django migrations.
PythonDjangoCeleryRediscareer
notes
Field notes on cost observability for AI agents: what to store, what to alert on, and why raw tokens beat computed costs.