Durable agents, a 12-week agent-built rewrite, and six Git fixes to roll out this week
Three AI picks: durable execution for agents that run for weeks, how Shopify rebuilt its app with agents in 12 weeks, and DeepMind's agent-driven formal verification of Signal's code. Three engineering picks: six Git vulnerabilities, AWS Nitro Sandboxes for per-pod VMs, and Dependabot alerts pausing on inactive repositories.
AI & Agent Engineering
Agent architectureverified
Durable execution for long-horizon agents: replay helps recovery, but retries still need idempotency keys
Kiran Prakash of Thoughtworks writes on martinfowler.com about agents that run for days or weeks. Durable execution records the result of each step that touches the outside world, so after a crash completed steps are skipped and state is rebuilt. Retries are at-least-once, so the article uses the purchase order number as an idempotency key, and treats the log as a published interface when the harness code changes.
Why it matters
Agents that wait on people and suppliers are distributed systems, with the usual failure modes plus non-determinism. The article is clear that durable is not the same as correct: lost outcomes, a world that changes during a wait, and code changes under running tasks still need design.
Do this week
Pick one agent tool that writes to an external system and check what happens if the response is lost and the call is retried. Add an idempotency key that stays the same across attempts before you add more retries.
Shopify rebuilt its mobile app in Swift and Kotlin in 12 weeks, with agents writing most of the code
About 30 engineers in pods of two moved the Shopify app, with more than 300 screens, from React Native to Swift and Kotlin in 12 weeks. Agents wrote most of the code. The team made screens runnable from the terminal without a device, shared one architecture across iOS and Android, and made CI fail when the platforms diverge. Written specs went stale, so the existing app served as the spec.
Why it matters
The gains came from the system around the agents: fast feedback, consistent patterns and checks in CI, not bigger prompts. Once code was cheap, review and validation became the bottleneck. This is a migration against a live reference; new features without one are harder.
Do this week
For your next agent-assisted migration, measure how long one feedback loop takes for an agent today. Cut that first, for example by making a unit of work testable without a device or full environment, and add a CI check for the pattern you most want kept.
DeepMind says AI agents formally verified security properties of Signal's protocol code in Lean
Google DeepMind describes AlphaProof Code, a multi-LLM agent framework for formal verification. Humans and models wrote the specifications, with experts reviewing each one, and agents generated over six million lines of Lean proofs about Signal's protocol code. Agents flagged about one percent for human review, and the work took four weeks. Two reintroduced past bugs were shown to be insecure. The artifacts are being open-sourced.
Why it matters
The split is the useful pattern: humans own what to prove, agents do the proving, and a checker verifies every proof. The four-week and first-of-its-kind claims are vendor-reported and depend on the specifications being right.
Do this week
Find one place where agent output is checked by a machine, not by eye, such as types, property tests or a schema. Ask whether a high-risk component could get a written spec that a checker enforces, and who would review that spec.
Git 2.56.1 fixes six vulnerabilities, including code execution during recursive clones
The Git project released fixes for six vulnerabilities in Git 2.56.1, with backports to 2.55.1, 2.54.1, 2.53.1, 2.52.1, 2.51.3 and 2.50.2. They include code execution when recursively cloning a malicious repository, path traversal through a server's advertised HEAD, memory corruption in merges and binary patches, and file access through fast-import streams. GitHub has patched github.com and lists GitHub Enterprise Server fix versions.
Why it matters
CI runners, build images and services that merge untrusted content often bundle their own Git and are updated less often than laptops. Hosting and merge services are named as affected by the merge bugs.
Do this week
This week, inventory where Git runs: developer machines, CI images, build agents and any service that clones or merges for users. Upgrade to 2.56.1 or the patched release for your series, and rebuild images that bundle Git.
AWS Nitro Sandboxes give each EKS pod or ECS task its own hardware-isolated VM
AWS introduced Nitro Sandboxes, lightweight VMs launched from an EC2 instance, integrated with EKS and ECS. Each sandbox has its own kernel and network stack, and AWS says neither cluster admins nor AWS can read pod or task memory. Network policies are enforced outside the sandbox. It runs on Nitro Enclaves instance types at no extra charge. Named uses include agent sessions and untrusted CI jobs.
Why it matters
A shared kernel is the weak point when containers run untrusted code. Per-workload VMs on existing nodes change the cost of a hard boundary. Performance and isolation claims are AWS's own, and the operational details still need testing.
Do this week
List workloads where you pay for a dedicated instance or node only for isolation, such as untrusted PR builds or tenant code. Pick one and test it in a sandbox, comparing cost, start-up time and debugging access.
From December 8, Dependabot alerts turn off on most repositories with no activity in a year
Starting December 8, 2026, GitHub will turn off Dependabot alerts on most repositories with no activity in the past year. Existing alerts are archived, not deleted. A push, creating the repository or turning alerts on counts as activity, and any push turns alerts back on. Repositories under an enforced security configuration that enables alerts are skipped. GitHub Enterprise Server is not affected.
Why it matters
Inactive repositories can still be deployed or depended on. If nobody pushes to them, their vulnerability alerts will quietly stop, which matters for libraries and services that are stable rather than abandoned.
Do this week
Before December 8, list repositories with no push in a year that are still deployed or used as dependencies. Put them under an enforced security configuration that enables Dependabot alerts, or archive the ones you no longer need.
Pick one agent or workflow step that writes to another system, such as creating a ticket or an order. In a test environment, drop the response after the write succeeds and let the retry run. Then add an idempotency key that stays the same on every attempt and run it again.
Why. It shows whether your retries can create duplicates today, which is the failure the durable execution article warns about.
Measure. Number of duplicate records created per forced retry, before and after the key. The goal is zero duplicates with the key in place.
Classic / essential read
Idempotent Receiver
Unmesh Joshi · 2 min
A client that gets no response cannot tell a lost request from a lost response, so it retries. This short pattern from Patterns of Distributed Systems is the base that today's durable execution article builds on.
One lesson. Retries are only safe when the receiver can recognise a repeat. Give each request a stable identity and return the saved result for duplicates.
Agents and infrastructure both reward boring safety work. Make retries idempotent, keep feedback loops short, patch the Git you forgot you bundle, and check which alerts will stop on their own.
Get this every morning.
Free. Same issue for everyone. One click to unsubscribe.