|4 min read
What Belongs in Agent Memory
A durable handoff separates observations, hypotheses and unfinished work, then records what would make each observation stale.
[AI & Data][Software Engineering]
18 posts
A durable handoff separates observations, hypotheses and unfinished work, then records what would make each observation stale.
The $7.99 rebuild bought a first version; the repository now shows the less photogenic work of preserving an archive and keeping its promises.
A review bundle should connect the requirement to evidence for the exact candidate and make the remaining decision explicit.
A CI alert becomes useful work when its revision, evidence and next owner survive the handoff; another prompt template does not supply them.
A useful agent acts within delegated authority, prepares consequential actions for review, and asks again when the action changes.
One paper shows frontier models degrade as context grows — even on trivial tasks. The other shows reasoning models hit a wall and think less as problems get harder. Read carefully, both point at the same engineering response.
Anthropic let Claude run a real shop for a month. It sold metal cubes at a loss, invented a Venmo account, and claimed to wear a blazer. The 'AI department that works while you sleep' is a genre — here's where it actually breaks.
Principles 26–30. The calibration layer that catches what the rest of the framework would miss: a PR-noise budget, independent verification, model-swap regression discipline, the 15-tool-call rule, and protecting junior development.
Principles 21–25. The governance and safety layer: strictKnownMarketplaces, no goal-conflict prompts, quarterly AppSec, four telemetry signals, monthly incident discipline.
Principles 15–20. The harness configuration that keeps the kernel and lifecycle cheap: CLAUDE.md under 200 lines, hooks for real incidents, skills that auto-invoke, subagent isolation, pinning, and Stage 5 distribution.
Principles 6–14. How work moves through an agentic engineering team: the ticket as contract, AI distillation with human curation, three gates, verification before done, characterisation tests, the 1.2× capacity rule, the J-curve, and telemetry.
Principles 1–5. The five rules that everything else in the framework rests on: standardise the harness, make verification load-bearing, default to plan mode, pick the cheapest layer, reflect every task.
Karpathy named one mode. Willison named the other. Most 'AI failed in production' stories are actually 'we promoted a vibe-coded prototype without transitioning into the production discipline.'
ReAct gave us a three-step loop. Production hardened it into five. The two new steps — Plan and Verify — are where everything that goes wrong, goes wrong. And the field has now named the worst offender.
Three open-source extractions converged on the same five layers. The architecture isn't a vendor narrative — it's a discovered structure. Here's the decision rule that keeps you from over-engineering it.
Most teams plateau at Stage 2 because they confuse 'we built skills' with 'we have a working AI engineering culture.' Here's the 5-stage diagnostic — and the moves that get you from Individual to Distributed.
The prototype-to-production gap for AI agents isn't technical — it's governance. Most organisations have nothing in this layer. The companies that build it first win the enterprise market. Everyone else stays in pilot purgatory.
When I first built Gluon on my Mac mini, I was solving a personal problem: monitoring Claude agents without losing my mind to tmux logs. But when teams join the picture, everything changes — security, governance, observability, and the fundamental role of the developer. Here's what production infrastructure for autonomous agents looks like.