
Use an agent only when the task needs dynamic decisions
Compare a fixed invoice workflow with a repository investigation to decide where dynamic agent behavior earns its execution cost and complexity.
The journal / Ideas in progress
Engineering & everyday learningSoftware, systems, and the things I’m learning along the way. From first principles to AI in production.
Find something to readA little theory. A lot of practice.

Compare a fixed invoice workflow with a repository investigation to decide where dynamic agent behavior earns its execution cost and complexity.

Derive central-difference error, expose floating-point cancellation and a hidden gradient bug, and test the limits of directional checks.

Review a Claude Code API patch for validation, authorization, response semantics, and test independence before accepting the generated diff.

Plan sorted runs and buffered merges for a file larger than RAM, calculate merge passes and I/O volume, and distinguish a byte budget from object memory.

Separate upstream attribution, a specific maintenance change and verified results through Capturia notices and a bounded GIF timing example.

Compare inherited descriptors with independent open calls using a six-byte file, and explain shared offsets, close behavior and close-on-exec.

Trace an alias to a local HTTPS server, observe TLS SNI and HTTP Host separately, and distinguish a certificate rejection from an application routing failure.

Give an agent explicit action, time, and cost limits, preserve enough budget for a useful handoff, and distinguish verified completion from exhausted effort.

Trace six independent instructions through a declared five-stage pipeline, calculate fill and drain overhead, and distinguish bubbles from global stalls.

Model a single FIFO help desk with explicit arrival ties, start times, waiting times, and idle periods, then separate service order from event scheduling.

Preserve the focused control across theme changes, retain a pointer fallback, and verify clipping, viewport clearance and forced-color behavior.

Trace hits, recency changes and evictions in a three-entry Python cache, then check loader failures, cached None and a scan that misses every time.