Ottto engineering notes

What we learn by operating coding agents.

Measured field notes on agent operations, context, cost, and reliability—drawn from systems we build and run.

2 published field notes Evidence, mechanisms, limits

Newest field note

Start with the evidence.

01 / 02
AgentsObservability

We audited our coding agent's memory. Drift was the real risk.

Across two Macs, 15 notes, and 66 local sessions, Claude Code auto-memory was a small share of estimated tool-originated context. Knowledge drift mattered more.

8 min read Read the field note

Earlier field notes

More from the workbench.

Prior investigations into the systems around coding agents.

02
11 min read

How We Built a Landing System for Agent-Written Code

One reasoning owner, compatible batches, and mechanical authority checks moved 17 agent-written PRs through two integrations in about two hours.

AgentsInfrastructure

What we investigate

Operational questions, tested in practice.

01

Agent operations

How autonomous loops are scheduled, handed off, supervised, and stopped.

02

Context and memory

What enters context, what persists across sessions, and where local notes drift from repository truth.

03

Cost and capacity

Where tokens, caches, subscriptions, and repeated pipeline work turn into spend.

04

Reliability and controls

Which failures recur, which safeguards hold, and what evidence earns automation trust.