The Voyager · Deep Dives

The Voyager · Deep DivesDeep Dives: one topic a day, explained properly.https://thevoyager.dev/🚀 A trillion-parameter open model lands. Your eval set decides whether it's yours to usehttps://thevoyager.dev/deep-dives/models-agentic/2026-11-02/https://thevoyager.dev/deep-dives/models-agentic/2026-11-02/Mistral Large 4's open weights (1T total, about 52B active, 1M context) arrive with mostly self-reported benchmarks; the week's building block is the eval set, calibrated judge and trace that turn those claims into a go/no-go.Mon, 02 Nov 2026 00:00:00 GMTmodels-agenticevalsllm-judgeobservabilityopentelemetrymistralopen-weights🧪 Lab, Step 1 (part 4): pull one piece out of a small GPT and see what you losehttps://thevoyager.dev/deep-dives/lab/2026-11-01/https://thevoyager.dev/deep-dives/lab/2026-11-01/Part 4 of Step 1: ablate a tiny GPT one component at a time, measure the damage against a seed-noise baseline, and learn which pieces are load-bearing and which are decoration at this size.Sun, 01 Nov 2026 00:00:00 GMTlablabtransformersnanogptablationresearch-ladder🛠️ Build: a retrieval harness that tells you what each choice is worth, or a contract-aware chunkerhttps://thevoyager.dev/deep-dives/build/2026-10-31/https://thevoyager.dev/deep-dives/build/2026-10-31/Two three-hour builds from this week's retrieval theme: a harness that scores chunking, context lines and fusion on a labelled set (runnable, numpy only), or a clause-aware chunker for contracts.Sat, 31 Oct 2026 00:00:00 GMTbuildbuildragretrievalchunkinghybrid-searchevalspython💻 claude-mem: retrieval for an agent's own memory, in three calls that each cost less than the lasthttps://thevoyager.dev/deep-dives/github/2026-10-30/https://thevoyager.dev/deep-dives/github/2026-10-30/thedotmack/claude-mem stores what a coding agent did, then serves it back through search, timeline and get_observations; the interesting part is the retrieval contract and its size caps, which you can copy for any RAG system.Fri, 30 Oct 2026 00:00:00 GMTgithubragagentsmemoryretrievalprogressive-disclosuretooling🎯 Principal/Staff Architect, project deep-dive: your retrieval stopped improving. Walk the panel through whyhttps://thevoyager.dev/deep-dives/interview/2026-10-29/https://thevoyager.dev/deep-dives/interview/2026-10-29/A practice project deep-dive in the style of a Meta Principal/Staff Architect loop: a RAG system you owned has plateaued, so how do you separate retrieval misses from insufficient context, and why can better recall make answers worse?Thu, 29 Oct 2026 00:00:00 GMTinterviewinterviewragretrievalevalsabstentionmetaproject-deep-dive🏗️ A knowledge assistant over policies and contracts: the retrieval design a professional-services firm can defendhttps://thevoyager.dev/deep-dives/system/2026-10-28/https://thevoyager.dev/deep-dives/system/2026-10-28/An end-to-end design for a policy and contract assistant at a professional-services firm: metadata-first retrieval, hybrid search with a reranker, per-document access control, and the numbers and threats that decide whether it ships.Wed, 28 Oct 2026 00:00:00 GMTsystemragretrievalhybrid-searchmetadata-filteringaccess-controlcontractssystem-design🔬 Three RAG upgrades that should add 47% added 17%: a 450-configuration study of what composeshttps://thevoyager.dev/deep-dives/research/2026-10-27/https://thevoyager.dev/deep-dives/research/2026-10-27/A 21 October paper stacks RAG techniques that each help alone: the top three retrieval gains predicted +46.6% NDCG@10 and delivered +17.2%, and the best pipeline changed as the corpus grew from 5K to 200K documents.Tue, 27 Oct 2026 00:00:00 GMTresearchragretrievalrerankingevalsablationsscaling🚀 Your citations code may 400 this week. Plus: how to design retrieval you can defendhttps://thevoyager.dev/deep-dives/models-agentic/2026-10-26/https://thevoyager.dev/deep-dives/models-agentic/2026-10-26/Anthropic's 2026-10-21 API version rejects citations on user-message text blocks, which hits RAG code that replays retrieved chunks; and the retrieval stack worth defending is hybrid search, contextual chunks and a reranker, measured on your own labelled set.Mon, 26 Oct 2026 00:00:00 GMTmodels-agenticragretrievalhybrid-searchrerankingchunkingclaude-api🧪 Lab, Step 1 (part 3): scale your tiny GPT without fooling yourselfhttps://thevoyager.dev/deep-dives/lab/2026-10-25/https://thevoyager.dev/deep-dives/lab/2026-10-25/Part 3 of Step 1: swap characters for a BPE tokeniser, grow the model in steps, and read loss against compute using the Chinchilla and Kaplan scaling papers as the yardstick.Sun, 25 Oct 2026 00:00:00 GMTlablabtransformersnanogptscaling-lawsbperesearch-ladder🛠️ Build: a tool gateway that says no, or a skill router you can measurehttps://thevoyager.dev/deep-dives/build/2026-10-24/https://thevoyager.dev/deep-dives/build/2026-10-24/Two three-hour builds from this week's ideas: a small MCP-style gateway with per-identity tool listing, deny-by-default calls and an audit log, or a skill-preselection harness that measures how routing recall falls as the catalogue grows.Sat, 24 Oct 2026 00:00:00 GMTbuildbuildmcptool-gatewayskill-routingauthorizationpython💻 context-mode: keep tool output out of the context window, and make the model's memory survive compactionhttps://thevoyager.dev/deep-dives/github/2026-10-23/https://thevoyager.dev/deep-dives/github/2026-10-23/mksglu/context-mode sits between an agent and its noisy tools: hooks reroute big outputs into a sandbox and a SQLite FTS5 index, so only an answer reaches the context window; the design is worth studying even if you never install it.Fri, 23 Oct 2026 00:00:00 GMTgithubagentsmcpcontext-engineeringtoolingretrievalsqlite🎯 Senior AI/ML Engineer, customer scenario: 140 tools, one agent, and a customer asking for more autonomyhttps://thevoyager.dev/deep-dives/interview/2026-10-22/https://thevoyager.dev/deep-dives/interview/2026-10-22/A practice question in the style of an Anthropic Senior AI/ML Engineer loop: a customer's agent has 140 tools across five MCP servers and is getting worse as it grows. Do you add autonomy, or take some away?Thu, 22 Oct 2026 00:00:00 GMTinterviewinterviewagentsmcptool-searchworkflows-vs-agentsanthropiccustomer-scenario🏗️ One MCP URL for a hundred engineers: what a tool gateway actually buys youhttps://thevoyager.dev/deep-dives/system/2026-10-21/https://thevoyager.dev/deep-dives/system/2026-10-21/How Eon replaced ten local MCP servers per developer with one authenticated gateway, and the three design choices that decide whether that pattern holds up: identity, policy and tool listing.Wed, 21 Oct 2026 00:00:00 GMTsystemmcpgatewayagent-identitycedarbedrock-agentcoresystem-design🔬 Skill routing collapse: your agent has 1,000 skills installed and loads half of nothinghttps://thevoyager.dev/deep-dives/research/2026-10-20/https://thevoyager.dev/deep-dives/research/2026-10-20/A 13 October paper shows that with 1,000 irrelevant skills installed, agents' recall of the skills they should load falls from 73.6% to 27.4%, and explicit preselection fixes much of it.Tue, 20 Oct 2026 00:00:00 GMTresearchagentsskillscontext-engineeringroutingevalstool-selection🚀 The agent that writes its own workflow: when the plan becomes a programhttps://thevoyager.dev/deep-dives/models-agentic/2026-10-19/https://thevoyager.dev/deep-dives/models-agentic/2026-10-19/Managed Agents can now let the model write a program that fans out up to 1,000 agents, which blurs the workflow-vs-agent line and moves the interesting design question to who approves the plan.Mon, 19 Oct 2026 00:00:00 GMTmodels-agenticagentsworkflowsorchestrationmanaged-agentsmulti-agentclaude-api🧪 Lab, Step 1 (part 2): the training loop is where small GPTs quietly go wronghttps://thevoyager.dev/deep-dives/lab/2026-10-18/https://thevoyager.dev/deep-dives/lab/2026-10-18/Part 2 of Step 1: read nanoGPT's train.py line by line, then rebuild its warmup, cosine decay, weight-decay groups, gradient clipping and eval loop around your own tiny GPT, and break each one on purpose.Sun, 18 Oct 2026 00:00:00 GMTlablabtransformersnanogpttraining-loopadamwresearch-ladder🛠️ Build: a fork lab that counts your agent's rescues and harms, or a session log that survives zombie workershttps://thevoyager.dev/deep-dives/build/2026-10-17/https://thevoyager.dev/deep-dives/build/2026-10-17/Two three-hour builds from this week's ideas: a paired-counterfactual harness that sorts every recovery step into rescue, harm or unchanged, or an offset-addressed session log with idempotent producers and epoch fencing.Sat, 17 Oct 2026 00:00:00 GMTbuildbuildagent-evalscounterfactualsdurable-executionevent-logpython💻 Durable Streams: the agent session as an HTTP log you can resume from any offsethttps://thevoyager.dev/deep-dives/github/2026-10-16/https://thevoyager.dev/deep-dives/github/2026-10-16/durable-streams/durable-streams is a small HTTP protocol for append-only, offset-addressed streams; the interesting design choice is treating the agent's token and tool-event stream as a resumable log instead of a socket.Fri, 16 Oct 2026 00:00:00 GMTgithubagentsstreamingprotocolsdurabilityhttpinfrastructure🎯 Forward Deployed Engineer, LLM system design: one customer, three latency budgets, one billhttps://thevoyager.dev/deep-dives/interview/2026-10-15/https://thevoyager.dev/deep-dives/interview/2026-10-15/A practice question in the style of an OpenAI Forward Deployed Engineer loop: design the inference tiers, cache layout and eval gate for a customer with live, near-live and bulk workloads.Thu, 15 Oct 2026 00:00:00 GMTinterviewinterviewsystem-designforward-deployed-engineeropenaicost-latencyprompt-cachingevals🏗️ An agent that survives a deploy: durable execution, and the history limit nobody mentionshttps://thevoyager.dev/deep-dives/system/2026-10-14/https://thevoyager.dev/deep-dives/system/2026-10-14/Why long-running agents fit workflow engines like Temporal, where the model call belongs, and the event-history ceiling that shapes how you design the loop.Wed, 14 Oct 2026 00:00:00 GMTsystemagentsdurable-executiontemporalreliabilitysystem-designopenai-agents-sdk🔬 Your agent's recovery step helps some runs and wrecks others. A paper shows how to measure thathttps://thevoyager.dev/deep-dives/research/2026-10-13/https://thevoyager.dev/deep-dives/research/2026-10-13/A 30 September paper treats agent error recovery as a per-run causal decision: fork from the same state with and without a refresh, count rescues and harms, and route recovery only where it pays.Tue, 13 Oct 2026 00:00:00 GMTresearchagentsevalsharnesserror-recoverycausal-inferencereliability🚀 The newest models won't let you force a tool call. Here's what your agent loop does insteadhttps://thevoyager.dev/deep-dives/models-agentic/2026-10-12/https://thevoyager.dev/deep-dives/models-agentic/2026-10-12/Opus 5.5 and Sonnet 5.5 return a 400 on tool_choice any/tool; the replacement is auto plus strict tool use, which changes how you design the agent loop's guarantees.Mon, 12 Oct 2026 00:00:00 GMTmodels-agenticagentstool-useagent-loopstructured-outputsclaude-apireliability🧪 Lab, Step 1 (part 1): build a GPT you can read in one sittinghttps://thevoyager.dev/deep-dives/lab/2026-10-11/https://thevoyager.dev/deep-dives/lab/2026-10-11/Step 1 of a public research ladder: grow a character-level GPT from a bigram table to a small transformer, watching the loss at each rung, with nanoGPT as the reference.Sun, 11 Oct 2026 00:00:00 GMTlablabtransformersnanogptattentionfrom-scratchresearch-ladder🛠️ Build: a prefix-aware context broker, or a task runner that can pause mid-thoughthttps://thevoyager.dev/deep-dives/build/2026-10-10/https://thevoyager.dev/deep-dives/build/2026-10-10/Two three-hour builds that turn this week's ideas (cache-aware routing, query-time context selection, suspendable agent tasks) into small Python systems you can run without a GPU.Sat, 10 Oct 2026 00:00:00 GMTbuildbuildkv-cacheroutingcontext-engineeringdurable-executionpython💻 Google's AX: Kubernetes-style YAML for agents, with Redis where etcd would chokehttps://thevoyager.dev/deep-dives/github/2026-10-09/https://thevoyager.dev/deep-dives/github/2026-10-09/google/ax declares agent tasks, workspaces and models as manifests and runs them in sandboxes with suspend/resume; the interesting part is the decision to keep short-lived tasks out of etcd.Fri, 09 Oct 2026 00:00:00 GMTgithubagentsruntimekubernetesdurable-executionsandboxinggoogle🎯 Interview, AI/Solution Architect (Google-style): design an enterprise assistant for 50,000 employeeshttps://thevoyager.dev/deep-dives/interview/2026-10-08/https://thevoyager.dev/deep-dives/interview/2026-10-08/A system design round for an AI/Solution Architect role: an internal knowledge assistant on Google Cloud, with a framework, a model answer, and the trade-offs interviewers actually probe.Thu, 08 Oct 2026 00:00:00 GMTinterviewinterviewsystem-designragenterprisesolution-architectgoogle-cloud🏗️ Your load balancer is quietly deleting your prefix cachehttps://thevoyager.dev/deep-dives/system/2026-10-07/https://thevoyager.dev/deep-dives/system/2026-10-07/Round-robin across vLLM replicas turns one warm prefix cache into eight cold ones; llm-d's precise prefix-cache routing is a good case study in fixing that, and in what its benchmark does and doesn't prove.Wed, 07 Oct 2026 00:00:00 GMTsysteminferencekv-cacheroutingllm-dkubernetescost-latency🔬 ReCAP: compacting agent memory from attention you already paid forhttps://thevoyager.dev/deep-dives/research/2026-10-06/https://thevoyager.dev/deep-dives/research/2026-10-06/A 30 September paper keeps the attention an agent computes anyway as a tiny graph, then picks context for the next request with no model call, cutting compaction latency by about 95% against Codex-style summarization.Tue, 06 Oct 2026 00:00:00 GMTresearchagentscontext-engineeringmemorykv-cacheattentioncoding-agents