Field guide / sub-domain
39
sources in this field
Updated September 20, 2026
Current thesis
The shortest path to orientation.
This synthesis records claims and practices from the cited sources; reported outcomes and product capabilities have not been independently verified.
Coordinator agents excel at batching work into 3–5 parallel streams with cross-model review (max 5 rounds) and adversarial verification. Production systems require versioned memory, concurrency control, five-stage loops with objective pass/fail verification. Task-level routing yields ~30% savings; heavy front-loaded planning prevents downstream chaos. Cold outbound GTM decomposes into four persistent agents (list builder, copywriter, campaign manager, infrastructure manager), each owning one artifact. Reliable multi-agent handoffs rest on three mechanisms: naming as join key, task-state-as-trigger (completion fires webhook waking next bot), 'blocking out loud' (agents post missing input and stop rather than guess). Claude Code as orchestrator with tools like ColdIQ MCP cut campaign builds from two days to one prompt and ~20 minutes. Ars Umbris functions as meta-harness, connecting multiple agent harnesses via MCP adapters to one durable workspace state. Parallelizing adversarial and browser-driven test runs across many instances makes exhaustive pre-release stress testing economically viable even for small teams, eliminating the tradeoff between thoroughness and resource constraints.
Evidence board
Claims worth carrying forward01A browser-based adversarial testing suite runs massively parallel adversarial checks against each new release, attempting to break it before ship, at a cost described as 'pennies.'
02Parallelizing adversarial/browser-driven test runs across many instances makes exhaustive pre-release stress testing economically viable even for small teams.
03Ars Umbris treats knowledge bases like codebases: typed files with explicit relationships and conventions an agent can parse, rather than freeform notes, so structure compounds across sessions instead of decaying.
04A repo bundles six parts as local files: knowledge (notes/sources), types (definitions), skills (agent instructions), MCP tools (operations), projections (UI components), and agent instructions/profiles—all composable and git-trackable.
05Repos declare dependencies via a manifest (repo.yaml with a deps list), similar to package imports in code, letting one repo (e.g. au-tree-research) supply vocabulary, skills, and MCP tools that other repos reuse without duplication.
06Cross-repo references use explicit syntax: [[note::another-repo]] for wikilinks and source::au-base-types* for shared types, making dependencies explicit like imports—so a workspace of multiple repos reads as one typed graph.
07Core thesis: knowledge work needs its own type engine (analogous to compilers/linters in software) because agents can silently accumulate broken links, missing fields, and inconsistent structures across sessions—reminding the model of conventions isn't sufficient at scale; independent checks are required.
08Type definitions (YAML) specify required fields and valid reference targets (e.g., a claim type requiring a source::au-base-types field); when an agent writes a note missing that field, the engine flags a diagnostic the agent can read via tools and fix—though judging whether a source actually supports a claim still requires human/agent judgment.
Adjacent fields
Latest evidence
Rafal Wilinski
Rafal Wilinski credits @typesafeai's Jev with building a massively parallel, browser-based adversarial testing suite that stress-tests each release automatically, at very low cost.
Heinrich
Heinrich launches the early alpha of Ars Umbris, a malleable IDE that treats a knowledge base like a codebase—typed files, composable repos, cross-repo dependencies, and a type engine that gives agents compiler-like feedback on knowledge structure.
Levi Munneke
A cold email agency architecture uses four persistent GrokBot agents (list builder, copywriter, campaign manager, infra manager) coordinated via a shared thread, with GPT-6 Astra as the coding/data layer, to run cold email GTM as a chain of SOP-defined artifacts with strict handoff rules.
Michel Lieben
A GTM agency's detailed playbook for running outbound campaigns through Claude Code as the orchestrator, with Apollo/Clay/Instantly-equivalent tools (via MCP) as endpoints. Covers folder structure, MCP connection setup, a seven-step signal-to-meeting workflow, skills files, reusable prompts, and eight recurring failure modes.
Y Combinator
YC convened researchers and founders to examine how agent 'harnesses' (scaffolding, context management, tooling) drive dramatically different performance from identical model weights, covering harness history, self-improving harnesses, and YC's internal agent stack (QM).
Shann³
A promotional post introducing GPT-6 Astra with a benchmark claim and three concrete workflows for using it: auditing agent session history for automatable patterns, auditing a personal knowledge base/second brain, and reviewing stalled 'evergreen' projects with a plan-then-delegate execution model.
Codez
A SpaceXAI engineer's tutorial on Grok Bot, a desktop app giving each agent its own cloud computer to log into tools, execute multi-tool workflows, and coordinate with other agents—shifting the user from prompting to delegating and managing approval boundaries.
Shahaf Antwarg
A workflow synthesis from a developer running 20 parallel coding agent sessions: role shift to tech-lead, isolated worktrees per task, five-part first-message discipline, early direction-checking over end-diff review, manual commits, and a living instructions file.
dex
Dex introduces /show-me, a skill for making coding agents communicate via compact visuals (component trees, call stacks, diagrams, pseudocode) instead of prose walls, then argues in a companion essay that 'lights-off' AI software factories fail because models can't maintain codebase quality over time, making human-in-the-loop program design and vertical slicing essential.
Adrian
Adrian outlines a structured planning workflow using Fable to front-load work before large-scale unattended agent execution: idea capture, adversarial questioning, research subagents, PRD drafting, policy setting, task decomposition into parallelizable 'beads', and adversarial review of tasks against the PRD.