The tooling landscape

Read this before the tables#

This is the most perishable chapter in the research. Tooling turns over quarterly, and several capabilities listed here as third-party tools became harness defaults during 2026, deferred tool definitions most clearly P.

Three rules for using it:

1. Verdicts are about the category, not the specific project. "Adopt a symbol-level retrieval tool" is durable advice. "Adopt this particular one" is not. Treat named projects as examples of a category to search for when you read this.
2. Verify before installing. These details come from project documentation and secondary coverage, not from running every tool. Anything you install runs with access to your code, and MCP servers receive your codebase content by design. Audit them like any dependency.
3. Check whether your harness already does it. The highest-value action here is often installing nothing. A tool that duplicates a built-in capability costs definition tokens to give you what you already had.

Verdicts use four levels: Adopt (use by default), Trial (worth a measured try), Watch (promising, unproven here), Skip (not for coding agents).

Layer 0: Prevention#

Layer 1: Retrieval#

Recommended stack: rg, a symbol-level MCP, and ignore files. Add semantic search only after measuring that concept questions are a real share of your work, and point it at docs/ and ADRs rather than code.

Layer 2: Tool surface#

The MCP ecosystem optimises for capability exposure. Context efficiency is capability restriction. Nothing in the protocol will restrict for you; the 30-minute audit is the tool.

Layer 3: State, memory and offload#

The most crowded category and the thinnest evidence. Read the memory section of chapter 6 before adopting anything here.

The gate for every tool here: if you cannot report retrievals that changed an action, per week after one month, delete it.

External knowledge stores#

Layer 4: Compaction#

Layer 5: Observability#

You cannot manage what you cannot see, and this is the thinnest layer.

Standards and conventions#

A starter stack, in order#

For a team using a CLI coding agent on a large repository. Each week is valuable on its own, so stop when the stop rule fires.

Week 1  prevention and deletion (mostly reversible; verify instruction changes)
        ignore files
        four shell wrappers for your loudest commands
        delete zero-call MCP servers only after a representative sample; verify task coverage and keep rollback
        instruction file → inference test → ≤ 150 lines
        a usage reporter + the transcript-grep script

Week 2  retrieval
        ripgrep by default, with -C 8 and exclusions
        a symbol-level MCP (Serena or your stack's equivalent)

Week 3  state
        a plan-file convention
        a .agent/ scratch directory with self-describing stubs

Week 4  boundaries
        semantic compaction triggers; one compaction per session
        a handoff template; mechanical session caps
        a live context display

Later, only if measurement demands it
        code execution against an API (a candidate for the largest remaining win; measure first)
        semantic search over docs and ADRs
        a memory system, with the retrieval gate from day one

Weeks 1 and 2 carry most of the value and need almost no ongoing discipline. Weeks 3 and 4 need behaviour change, which is where adoption usually stalls.

What the ecosystem is missing#

These are gaps a builder could fill. Their absence explains several of the field's blind spots.

1. Cache-aware accounting. Nothing widely available reports cache-adjusted cost, so nobody sees regressions like the dynamic-tool case.
2. Post-boundary instrumentation. The most error-prone moment in a session is unmonitored everywhere by default.
3. Pass^k runners for local workloads. Running your own tasks k times and computing Pass² should be one command. It is not.
4. Read-coverage auditing. Separating starvation from dilution still means reading transcripts by hand.
5. Packaged addressable-recall compaction. The strongest published compaction result has no widely available implementation.