01 Pressure
Goal arrives
Intent, files, context, budget, and risk enter the same loop.
scopeAI-native systems engineering
demo clean / production messy
Notes on building systems that hold up in production.
Notes on retries, state, budgets, coordination, observability, and failure recovery — the details that decide whether pressure becomes progress or repeat pain.
distributed systems production agents data platforms observability
production-agent-loop runtime map
Core question Can the loop recover when tools, cost, and memory drift?
02 Control
Policy bounds
The runtime chooses tools, retries, memory, and stop conditions.
bounds03 Signal
Telemetry answers
Traces expose latency, token spend, tool errors, and bad branches.
p9504 Library
Learning captured
Failures become evals, docs, guardrails, and better defaults.
replay How an agent survives production pressure A production agent is not a chat box. It is a bounded control loop with state, tools, policy, observability, and replay.
10B+ events/day systems direction
150+ paying users from zero
70+ data pipelines shipped
1B+ market data points served
Search
Find the exact thread.
Jump straight into the essays, concepts, explainers, and deep dives behind the system you are thinking about.
agents
AI systems that survive real users
Multi-agent orchestration, bounded tool calls, memory, evals, cost guardrails, and recovery paths.
systems
Distributed systems with operational taste
Coordination, retries, idempotency, stream processing, failure detection, SLOs, and quiet incident prevention.
platforms
Data platforms that earn their keep
Kafka, Spark, ClickHouse, search, lakehouse migrations, and the product pressure behind architecture choices.
Latest field note
Continue reading
Blog
Recent posts
01 · Seven Checks for Systems That Look Correct Seven checks connecting refreshed data, Kafka and AI series. 02 · 95% Agreement, Zero Failures Caught: Testing Your LLM Judge A useful judge must catch the failures your release gate cares about. 03 · Your MCP Server Is an Authorization Boundary Carry identity and consent all the way to the target object. 04 · Your Cache Hit Rate Is Not Your Token Savings Explain the bill with token categories, not request hit rate.
Series
Deep dives
Harness Engineering: The Compounding Stack The operating layer around serious AI work. Harness Internals A concrete engineering pattern, stripped down to the decision that matters. Production Agents Deep Dive Retries, state, cost, security, and evals after the demo. AI Engineering Fundamentals Start at tokens; end at evaluated agents.