Stack

What I actually run

Not a leaderboard. This is the current loadout, with the honest verdict: what gets used every day, what is on trial, what I dropped and why. Revised when an experiment changes my mind - the revisions are the point.

Last revised · November 2026

Harnesses

4

The thing that turns a model into something that can actually edit a repo.

ToolRoleVerdict
DeepSeek HarnessLocal agent runtime; tools, subagents, workflows, permissions.Daily driverdailyEvery experiment in the lab runs here first. The permission model and the audit trail are the reason.
Claude CodeTerminal agent, strong on long refactors and repo-wide reasoning.Second opiniontrialUsed when a task needs a very long single-threaded session. Resumability is good; cost tracking still manual.
CursorIDE-native agent, fast inner loop on small diffs.EditordailyStill the best place to review and steer a diff line by line. I do not trust it with anything wide.
AiderGit-native pair programming, tiny surface area.FallbackwatchSurvives every tooling shakeout because it does one thing. Good for scripted, repeatable edits.

Verification

2

The part everyone skips and then pays for twice.

ToolRoleVerdict
Vitest + PlaywrightUnit/integration tests plus a real browser.Ground truthdailyIf the agent cannot make a test pass, it has not finished. Everything else is opinion.
mutation-style probesDeliberately break code, check the suite notices.Trust checktrialThe cheapest way to find tests the agent wrote to match its own bug.

Context & retrieval

3

What the agent is allowed to know, and at what price.

ToolRoleVerdict
Repo docs as specARCHITECTURE.md, ADRs, task files in-tree.PrimarydailyVersioned with the code, reviewed like code. Beats any external knowledge base I have tried.
MCP serversStructured tools over ad-hoc scraping.IntegrationdailyWorth it when the data is genuinely external. Not worth it as a way to avoid writing a script.
Vector search over the codebaseEmbeddings for code retrieval.DroppeddroppedConsistently worse than grep plus a good file map. Revisit when the repo exceeds a few million lines.

Orchestration

3

More than one agent, or one agent more than once.

ToolRoleVerdict
SubagentsScoped children with their own context.Daily driverdailyThe single biggest quality win so far: the parent stays clean, the child can be wrong safely.
Workflow scriptsFan out N tasks, collect structured results.Batch worktrialExcellent for audits and migrations across many files. Overkill for anything with one obvious path.
LangGraphGraph-based agent orchestration in Python.PrototypingwatchFine for research code. I have not kept a production system on it.

Observability

2

If you cannot see the run, you cannot improve it.

ToolRoleVerdict
Transcript logsFull prompt/tool/result trails, kept on disk.Non-negotiabledailyRaw JSONL per session, committed next to the experiment. Every post in /lab links to one.
Cost + latency ledgerToken spend and wall-clock per task.Decision inputdailyRecorded per task, not per month. A monthly bill tells you nothing about which workflow is worth keeping.