Overview

Portia is a genus of jumping spiders that hunts other spiders, prey dangerous enough to kill it if an approach goes wrong. It plans routes that take it out of sight of its target for up to an hour, adjusts its tactics per species, and backtracks when a route goes wrong. It does this with a brain widely cited at around 600,000 neurons—about the size of a pinhead, and a small fraction of what most predators of comparable behavioral complexity carry.

There is a pretty valuable lesson here, applicable across both the personal and professional spectrums: How even a minimal architecture can extract maximum value within a constrained budget. When there is no budget or headroom to just add capacity, every strategy and every choice has to earn its place.

When using AI, that is the same constraint agent orchestration runs into once a task grows past what a single model can handle well in one pass. The fix, as Mother Nature’s Portia shows, is not a bigger brain. It is learning to do more with less, and to do it well: Honing a small number of well-chosen capabilities, and wiring them together deliberately for maximum value.

The design constraint: No room to outmuscle the problem

Portia cannot outrun or outpower its prey. Much like us apes, Portia out-thinks it instead. Research on its pre-planned detour behavior documents routes where the spider loses direct sight of its target—but out of sight does not mean out of mind. It holds that target’s location in working memory and uses it to complete the approach correctly.

That is the same constraint behind agent orchestration decisions that every user runs up against eventually, even with the most lavish token budget. No one has infinite context or unlimited model calls; the architecture has to be deliberate about what it holds in memory, what it hands off, and when.

Three narrow capabilities, not one big one

Portia’s hunting success does not come from any single trait. It comes from three narrowly defined skills, working together:

  • Working memory of a hidden target. It tracks a representation of prey it can no longer see, for the length of a detour that can run close to an hour
  • Tactics tuned per target. It adjusts its approach and deceptive signaling by species rather than running one generalist strategy against everything it hunts
  • Trial-and-error with recovery. Research on Portia’s confinement problem-solving shows it learns from both successful and failed routes, and can recognize and correct a wrong turn instead of abandoning the approach

In isolation, none of these will achieve the goal. However, working together and time\-sharing on a 600,000-neuron compute platform, they produce behavior that researchers have compared to vertebrate-level planning.

The same pattern, different substrate

This maps directly onto my agent harnesses, such as those running on my homelab: An orchestrator decomposes a task, routes each piece to whichever specialized model handles it best, and writes intermediate state to persistent memory so nothing gets lost between steps.

  • Working memory of a hidden target. This translates to a skill defining a markdown file or a vector database such as Hindsight or Graphiti, and how to use it to track location
  • Tactics tuned per target. This skill defines a dictionary of MCP calls, specifying which services and methods are the last known best fit for a given target
  • Trial-and-error with recovery. This skill hands control off to the orchestrator, such as n8n or LangGraph, to maneuver the entire system from one state to another within a state machine, slowly but surely loading and unloading relevant models and pointing to relevant information between instantiations

Graph by Tyler Style

Neither system gets its result from one larger component doing everything. Both get it from holding state across steps that lose direct visibility into each other, and from routing each sub-problem to the narrowest tool built for it, running on tightly constrained compute resources.

The tradeoff: Precision costs speed

Researchers studying Portia note that despite its planning ability, it is a comparatively slow thinker. A single detour can take the better part of an hour. That is a tradeoff and not a side effect. Portia cannot observe the target directly during a detour, and holding that plan in working memory across that detour is what costs the time that a faster, more reactive predator would not spend.

The same tradeoff shows up in orchestrated agent systems. Decomposing a task, routing it to specialized components, and persisting state between steps adds real overhead compared to one model handling a task end to end in a single pass. That overhead is worth paying when a wrong call is expensive to walk back, which is exactly the case my AI-augmented refactor pipeline was built around. It is not worth it for a task that is small, bounded, and cheap to redo if it fails.

Takeaway

Portia’s advantage is neither brains nor brawn. It is clever architecture working within a fixed, small budget to hold state across actions that lose sight of each other, and route each piece of the problem to the narrowest capability built for it.

The same principle holds when the substrate is model calls instead of neurons. You can brute force a hard problem by throwing more resources at it: More parameters, a bigger context window, and more compute. However, returns diminish fast, and the expenses increase exponentially.

On the other hand, a well-designed system that routes work to the right-sized component and remembers what it is doing costs more in strategizing up front but is the cheaper and more reliable one in the long run.

The article originally appeared on LinkedIn.