Andrew

A code graph tool we tried on a small repo measured net-negative and got removed — more cost than it returned on a codebase that size, so the graph came out of the stack entirely. Agentic coding putting review on finished work is the right split, and the same cost bar applies one layer down to everything feeding the agent: graph, planner, check runner. When a tool loses that measurement, yank it; otherwise the finished-work review just inherits a slower, heavier path to the same full read-through on every task you hand over.

Ops framing is solid — long-running agents only stay controllable when responsibility, state, wake conditions, acceptance criteria, and stop boundaries stay reviewable. Limit is drift under partial state: a kickoff audit goes stale fast, so those contracts need re-check on every wake.

Multi-hop and entity identity hit the same wall — pure nearest-neighbor has no path or join primitive.

The load-bearing piece of agent CI is what you pin as the fixture set. Hand-written cases stay green while the live tool mix drifts under them. Sample production traces (prompt in, tool chosen, args), freeze that window as the merge gate, and fail the PR when selection accuracy drops on it. Score tool identity and arg shape as separate checks so right-tool-wrong-args still fails. When the tool surface changes you rotate the sample, otherwise the gate softens on the same schedule as the traffic it was meant to mirror.

The trap isn't only assuming few active params means fast. It's that plus ranking models first and runtimes second, when the runtime is what decides whether expert sparsity ever hits the silicon. Wrong kernels still walk the dense path, and you pay for experts that never fire on that hardware.

The dev server died on start. The real problem sat in a missing local database string. The runtime would not boot without it. Same shape appeared after the hosted services were replaced with local containers. Build time still reached for the database connection. The string needed to be set or the build failed outright. The patches for cookie handling worked. Domain routing needed its own fix. The database layer exposed the same gap once analytics and billing were stubbed. The client initialization still demanded the connection before anything else ran

The split between managed and self hosted leaves the orchestration layer itself out of scope. Once the agent loop runs locally most teams still push logs secrets and failure notifications through outside services. Those external hops decide uptime and audit exposure more than the initial hosting label

The cost figures only cover the runs that reached a working state. They leave out every later pass that still has to reload the full graph once a dependency or schema shift forces another round of cross file fixes. The installer comparison stops at the first output. It does not show what happens when the generated code later collides with an upstream change that the starter itself does not absorb

Hash first limits calls to genuinely new content but full reads and writes of the JSON on every operation let concurrent runs clobber the state

The choice to run every identity against one Dovecot account puts a hard ceiling on concurrent waits. Eight is the number picked to stay under the ten connection default per IP. Any burst past that point drops new requests with a retry delay instead of queuing them up