Which Species for Which Job? Mapping Agent Architectures to Real Work

Not all L3 agents are the same. Deploying the wrong architecture guarantees failure — regardless of the model or the prompt.

I learned this the hard way. Last year I tried to run a content pipeline through a coding harness — the same architecture that handles individual code tasks beautifully. It was a disaster. The harness was designed for a human reviewing one output at a time, not for processing 500 pieces of content in a batch. I was using the right autonomy level and the wrong species.

After studying production deployments across dozens of teams, researcher Nate Jones identified four distinct agent species. Each is built for a different kind of work. Two systems can both be L3 — fully autonomous, tool-using, iterative — and still fail if one is deployed where the other belongs.

---

Species 1: Coding Harness

The human acts as manager. The agent acts as engineer. You set the task, the agent produces output, you judge the work.

This is the most mature species. Boris Cherny says “coding is solved” — 27 PRs a day from his phone. That’s a coding harness at work. Peter Steinberger pushed this further with OpenClaw: running multiple harnesses in parallel, each on 20-minute cycles, with human review at output.

Best for: Individual contributor work, bounded coding tasks, anything where a human can judge quality at the output boundary.

Key trait: Single-threaded by default. Scaling means running multiple harnesses in parallel, not making one harness more complex.

---

Species 2: Dark Factory

A spec goes in. Software comes out. No human reviews intermediate outputs.

This is the assembly line model. The agent optimizes against your specifications, not against your attention. Human involvement at two points: writing the spec and reviewing the evals.

Best for: Well-specified, bounded, repetitive production work — automated code migration, bulk refactoring, content pipelines.

The risk is the 3% problem magnified. A dark factory producing 10,000 outputs at 97% correctness creates 300 failures no human reviewed. Acceptable only when failures are low-cost or evals catch them before production.

---

Species 3: Auto Research

Descended from classical machine learning. The agent hill-climbs against a metric — optimizing conversion rate, code quality, response accuracy, or any quantifiable target.

Auto research only works when success can be measured numerically. If you can’t define a metric, you’re not doing auto research.

Best for: A/B test optimization, prompt tuning, hyperparameter search, any problem with a clear score and a large search space.

Key trait: The agent iterates autonomously, trying variants and measuring results. No human judgment at the iteration level — just at the objective-setting level.

---

Species 4: Orchestration Framework

Multiple agents lined up with a coordination layer handing work between them. Writer → editor. Researcher → drafter. Planner → executor.

This is the most complex species and the most commonly over-deployed. Jones’s rule: “The art of building good agents is often the art of finding different simple configurations that enable the agent to do the particular work you have in front of you.”

Best for: Multi-step processes where different steps need different skills and volume justifies the coordination overhead.

Critical constraint: Scale-dependent. Worth it at 10,000+ tickets. Overkill at 100.

---

The Selection Cheat Sheet

---

The Common Thread

Species can hybridize. All three frontier model providers have converged their harnesses toward hybrid capabilities — Claude Code added coordinator mode, Codex CLI added gateway integrations, Gemini added agent runtime with tool orchestration. The direction is clear: harnesses are becoming species-agnostic at the feature level.

But every species needs the same foundation: a harness that’s replaceable, memory that’s portable, and an environment that’s owned by you. The species determines what the agent can do. The environment determines whether you can keep doing it when the vendor changes terms.

That’s the last problem in this series — and the most important one.

---