Posts
-
81% of my agent's telemetry was a copy of something I already had
My agent logged every tool call verbatim, and two thirds of the biggest rows were the harness echoing back a file that was already in git. Measuring it split one population into two with opposite economics, and produced an option nobody had drafted.
-
Prompt injection is worse when your agent has a memory
Most injection defense assumes the attack ends with the session. A memory system removes that assumption: one poisoned line promoted into durable memory resurfaces later as trusted context. So the defense has to be infrastructure, not vigilance.
-
An agent that remembers everything is a secret leak with a good memory
An encrypted vault protects the secrets I chose to protect. The dangerous ones are the keys pasted into a debugging session, and a memory system that writes every event to disk turns each one into a liability that grows every day the system runs.
-
Testing a data pipeline against the spreadsheets it replaced
Replacing a reporting workbook means first proving you match it. Then the diffs start pointing the other way, and you find out the report you were treating as the specification does not reconcile against the warehouse it came from.
-
Letting an agent write to my production CMS
Handing an agent write access to a live CMS is not a drafting problem, it is a blast-radius problem. The work that made it safe happened before the agent existed: collapsing eighteen write paths into one seam that a test can enforce.
-
Running parallel agent sessions without them stepping on each other
I run several coding-agent sessions at once, often on the same repo. Two failure modes make that miserable: corrupted shared state and duplicated work. One was solved by construction, the other by presence - and the fix deliberately contains no locks.
-
How I measure whether my agent's memory works
My coding agent could not recall a fact I had explicitly saved. The root cause was invisible, the obvious fix was wrong, and the measurement that proved the right one nearly lied to me twice.
-
My agent system dreams at night, and that is where its memory comes from
Durable memory does not come from an agent deciding mid-task what to remember. It comes from a nightly consolidation cycle: cluster the day's events, reflect with an LLM, score, gate hard, and promote very little. The gates are the whole design.
-
Documentation my agents can trust (and why it made me faster too)
My technical posts get written in an afternoon, weeks after the work, without reconstructing anything. That is not memory. It is a documentation standard designed for agents: immutable decisions, plans with stable task ids, and one authoritative home per question.
-
Giving my agent a map of the codebase (drawn partly from its own footsteps)
Every agent session used to start with grep archaeology: rebuilding a mental model of the repo from scratch. Now each project carries a code map fusing static structure, git history, and a signal most tools ignore - which files the agent itself touched together in real sessions.
-
Audition the voice before you commission the work
A persona config for AI-generated writing is write-only: you describe a tone in prose and find out whether it worked after paying for the full piece. The fix is a preview agent so cheap and so constrained that sampling the voice becomes free.