leadbybuild.ingArchive

LeadByBuilding Newsletter · 8 min read

How agents should remember, how to map a DR boundary, and what a vulnerability feed can and cannot tell you

Two AI picks: Anthropic on agent memory design and TUI's metadata-based context selection for threat modeling. Two engineering picks: AWS on running a DR readiness engagement and Cloudflare's Zero Wire vulnerability briefings.

AI & Agent Engineering

Agent memoryvendor-reported

Anthropic: let the model judge what to remember, and use offline passes to correct its memory notes

Anthropic describes agent memory as notes a model keeps between sessions. In its tests, telling Claude only that a memory folder exists worked for stronger models, while a rule to record only failures cut the gain from memory. Offline review passes called dreaming checked notes against transcripts and fixed a wrong note. A wiki built over 128 sessions raised a 25-question score from 13% to 48%.

Why it matters

Memory notes can be wrong from the start or go stale, and a later session will trust them. Rigid rules about what to save can hurt capable models. These are Anthropic's own evaluations, so treat the numbers as vendor-reported.

Do this week

If your agent has a memory directory, add one scheduled pass that re-checks notes against the session logs and corrects or deletes the wrong ones. Do not add write rules until you have measured their effect.

Anthropic →
Agentic workflowsvendor-reported

TUI cut threat-model prompt context by 94% by selecting standards from metadata instead of loading all of them

TUI built a LangGraph threat-modeling agent on Amazon Bedrock. Loading all 192 internal standards (338 KB) into the prompt failed. The model instead reads pre-generated metadata (about 60 KB) and picks 16 standards, which is 21 KB. A validation node requires all six STRIDE categories and one gap analysis before the workflow continues. TUI reports sessions rose from 3 to 4 a month to 10 to 12.

Why it matters

The useful pattern is selecting context from a bounded corpus rather than stuffing it, plus hard completion checks the agent cannot skip. Throughput and time figures are TUI's internal measurements, published by AWS.

Do this week

Look at one prompt in your own agent that carries a large rulebook. Try a metadata index, with keywords and applicability, that the model searches first, and compare output quality against the full dump.

AWS Architecture Blog →

Engineering & Leadership

Reliabilityverified

AWS: a five-step DR readiness engagement, and why a first failover missed a 5-hour target by more than double

AWS Part 3 of a disaster recovery series lays out five steps: kickoff, assessment, validation plan, execute and learn, and a quarterly cadence. It starts with a boundary map of every service and account in the recovery path, each with an owner. In its example, the first exercise missed a 5-hour commitment. The causes included a shared authentication platform outside scope and a manual DNS step that took over 3 hours.

Why it matters

Most recovery targets fail on dependencies nobody mapped, not on the app itself. The advice to pass a planned failover before a surprise drill avoids noisy findings.

Do this week

This week, list the shared services your top customer journey needs to recover, such as auth, DNS and secrets, and write down the owner and stated recovery time of each. Any gap in that list is your first DR action.

AWS Architecture Blog →
Security operationsverified

Cloudflare launches Zero Wire, a public feed of critical vulnerability briefings

Cloudflare introduced Zero Wire in Cloudflare Radar. Each briefing explains a critical vulnerability, who should investigate, recommended actions and links to evidence and vendor guidance. It separates Needs attention from Reference, and treats severity and active exploitation as distinct. Cloudflare also offers a prompt for a read-only agent assessment of your own deployments.

Why it matters

The format reflects a real triage problem: more disclosures and less time to decide if you are affected. Using such a feed is a convenience, not a substitute for your own asset inventory.

Do this week

Try the read-only assessment prompt on one recent briefing and see whether your agent can answer from your inventory which assets run the affected software. A failure to answer shows an inventory gap.

Cloudflare Blog →

Worth experimenting with

Add a memory audit pass to one agent

Pick an agent that keeps notes between sessions. Once a week, run a separate pass that re-reads recent transcripts and flags notes they contradict. Have it correct or delete those notes and log the change.

Why. Wrong notes get trusted in later sessions. A check against transcripts catches errors without limiting what the agent may write.

Measure. Count notes corrected per pass and the task failures traced to a stale note over two weeks. If both are near zero, drop the pass.

Classic / essential read

Monolith First

Martin Fowler · 4 min

Fowler notes that nearly all successful microservice systems he heard of began as a monolith, and that stable service boundaries are hard to draw early. It pairs with this edition's point that unmapped dependencies and boundaries cause failures.

One lesson. Before splitting a system, check whether you know the boundaries well enough, and keep the monolith modular so a later split stays possible.

Read the original →

Commute watch · optional

Context Engineering for Agents - Lance Martin, LangChainLatent Space · 63:18Surviving the K8spocalypse: The Untold Saga of Multi-Site Disaster Recovery - Nic VermandéCloud Native Days Italy · 31:01

Treat memory, context and recovery plans as things you verify. Check agent notes against logs, and check recovery plans against the owners and dependencies you can actually name.

Get this every morning.

Free. Same issue for everyone. One click to unsubscribe.

Subscribe