verification layer

Agent Memory Products Remember Your Conversation. Nobody Is Remembering Your Codebase

Every memory tool shipped this year stores what you said. Measured sessions of Claude Code auto compaction reduced 132,000 tokens of accumulated state to roughly 2,300, a 98% reduction, and what survives is the transcript rather than the understanding. Multiply that by 40 engineers separately paying a model to work out what the same billing service does, and the cost of not writing the answer down becomes visible.

A Team of Agents Built a System in a Week and Left €15M of Unmaintainable Code

The number that gets quoted from that case study is the token bill, somewhere between €10M and €15M for one week of autonomous agent work. The number that should get quoted is the second half of the sentence, which is that the resulting code was described as nearly unmaintainable. The bill was paid once. The code has to be lived with.

The Security Pass Rate on AI Generated Code Has Not Moved in a Year, and Better Models Did Not Help

Veracode has now evaluated more than 150 models and the average security pass rate sits at about 56%, against 55% in their first report. Across that same period coding benchmarks improved steadily, larger models did not outperform smaller ones, and vendor claims about security aware training did not correspond to measured outcomes. A capability that stays flat while adjacent capabilities improve is telling you the bottleneck is somewhere other than the model.

Waiting for a Bigger Context Window Is a Bet Against Four Orders of Magnitude

At a conservative 1,000 tokens per source file, 10 million files is roughly 10 billion tokens. A 1 million token window holds 0.01% of that, and a 10 million token window holds 0.1%. Closing that by scaling needs something like a 10,000x increase rather than a 10x one, and quality per token moves against you while you try.

Agent Skills Tell the Model How to Work. Nothing in That Loop Checks the Result

Skills are the best thing to happen to agent workflows in the last year, and we write them ourselves. This is not an argument against using them. It is an argument about what category of thing a skill is, because a large number of teams are now treating a code review skill as the answer to whether AI generated code is safe to merge, and a skill structurally cannot be that.

We Built claude-context's Architecture Ourselves in October 2025, and It Failed on Accuracy

claude-context is Zilliz's open source code search MCP server, around 11,800 stars, chunking along AST boundaries and serving hybrid BM25 and vector retrieval with Merkle tree incremental reindexing. The engineering is competent. We cannot credibly write a takedown of it, because we shipped the same architecture in October 2025, watched it fail on accuracy, and threw it away. This is what that cost us to learn.

Augment Code Measured Our Thesis for Us, Then Stopped One Step Short

When Augment exposed its context engine over MCP in February 2026, Claude Code on Opus 4.5 gained around 80% on quality and Cursor gained around 71%, with code completeness up 60% and correctness improved fivefold. Somebody else measured that, on somebody else's product, and it is the strongest evidence anyone has produced that context architecture matters more than model selection. Here is where our architecture diverges from theirs.

Your CLAUDE.md Was Measured, and It Was Roughly Net Negative

An ETH Zurich evaluation found machine generated context files lowered success rates and raised costs by roughly 20%, human written files bought about 4% on the harder benchmark, and Claude Code was the only agent tested where even developer written files failed to beat having no file at all. The most revealing result came from a control: strip the existing READMEs out and context files finally start helping.

Code Metal Proves Two Programs Are Equivalent. Nobody Asked What Either One Means

A $125M Series B into formal verification for defence is the strongest market proof anyone has produced that enterprises will pay real money to know whether code is correct. It also published its own ceiling, in public, and that ceiling is the reason we took a different route: equivalence is a relation between two artifacts, and it is silent about purpose.

A Call Graph Is Not a Specification, and CodeGraph Shows Exactly Why

CodeGraph went from launch on 18 January 2026 to roughly 47,000 stars in five months by mapping symbols and call edges locally, with no server and no cloud embedding call. It deserved every one of those stars. It also answers one class of question completely and another class not at all, and the second class is where production breaks.

The Diff Is the Wrong Unit of Review, and CodeRabbit Is Built on It

A pull request changes 11 lines, the tests pass, every automated reviewer approves, and three weeks later a reporting service is quietly returning wrong revenue. Nothing about the diff was incorrect. What was wrong lived in another repository nobody in the pipeline had read. This is where diff scoped review has nothing to say, and it is now the majority of what breaks in a large organisation.

GitHub Spec Kit Taught Millions of Developers the Right Idea and Left the Hard Half Undone

Spec Kit spent a year and gathered well over 100,000 stars teaching several million developers that the specification, not the code, should be the source of truth. That is category creation at a scale no startup could have afforded, and it lands directly on our thesis. It also assumes a blank page, which is true exactly once in the life of a system.

GitNexus Went From 1,200 Stars to 42,000 in Two Months, on a Design That Stops at One Machine

The deepest MCP integration anybody has shipped, with 16 tools, 7 resources, skills and Claude Code hooks, running with zero server on a local graph. The developer experience is the best in the category and that is why it grew. Two things are worth thinking through before you build anything organisational on top, and only one of them is technical.

Graphify Builds a Beautiful Graph of One Repository. Production Breaks Between Them

Local deterministic parsing over 36 tree sitter grammars, typed nodes, provenance tagged edges, Apache 2.0, no vector store and no telemetry. That is a clean design and the privacy position is better than most of the category. The problem starts the moment your organisation has more than one repository, and it is not a problem more parsing solves.

Greptile Indexes the Codebase, Which Is Necessary and Not Sufficient

Of every commercial tool in automated code review, Greptile made the closest architectural call to ours: index the whole repository first, then review the change against it. That decision is why it catches 82% of bugs against CodeRabbit's 44%. It also arrives with about 11 false positives per run, and the shape of those false positives tells you exactly what an index of symbols can and cannot know about intent.

Repomix Compresses Your Repository by 70%, Which Does Not Solve the Problem

Repomix is the leader in context packing, with 26,000 stars and 255,000 npm downloads a month, and its tree sitter compression genuinely cuts token count by about 70%. Apply that to 10 million files and you go from three orders of magnitude short of the context window to slightly less than three orders of magnitude short. Every tool in this category publishes a token reduction number. None of them publishes a correctness number.

Serena Knows Every Symbol in Your Codebase and None of the Meaning

Serena is the best open source symbol level tool for agents, with 25,200 stars, 170 contributors and over 40 languages behind a clean LSP to MCP design. A language server gives you exact call edges and zero intent. In the MSR 2026 dataset of rejected agentic pull requests, only 36% failed on the code itself and 31% failed on a decision the project had already taken, which is the third no symbol table can reach.

Sourcegraph Answers Where the Code Is. Agents Need to Know Why It Exists

Sourcegraph built the most complete cross repository code search anyone has shipped, and it did so on an assumption that held for fifteen years: a human is asking. Humans supply the meaning the index leaves out. An agent cannot, and at 10 million files a 1 million token window holds 0.01% of the estate, so the gap is not one model generation away.

Context Rot Is Not a Session Hygiene Problem. It Is an Architecture Problem

Compliance with an agent's own explicit constraints falls from 73% at turn 5 to 33% at turn 16 across 4,416 measured trials, and Claude Code auto compaction reduces 132,000 tokens of session state to roughly 2,300. Shorter sessions and cleaner instruction files treat the symptom. The window fills with low value tokens because the agent rebuilds an understanding of your codebase from raw source on every task.

Running an MCP Server Inside DeepSeek Harness: What the Bridge Carries, and When You Need a Plugin Instead

We pointed DeepSeek Harness at an MCP server we already ran and it worked without a line of new code. This is what the official bridge carries, what it quietly drops, where plugins get published, and the one reason a verification tool cannot live behind MCP at all: over MCP the model decides whether to call you, and a check the model can skip is not a check.

The Monorepo Versus Polyrepo Debate Came Back Because Agents Cannot See Across a Boundary

This argument was being compared to the Vim and Emacs wars in 2024, and then it returned everywhere inside six months. Something dragged it out of retirement, and it is not really about where files live. The variable that decides whether agents work is whether the question of what depends on this is answerable at all, and neither layout answers it.

Everyone Open Sourced the Hands This Year. Nobody Open Sourced the Check

Agent runtimes got very good very fast. The thing that decides whether a change is correct across your other repositories did not get built at all. On cross repository blast radius, why AI code review misses it, and what a check actually needs.