A capable model does not determine whether an agent finishes a long task. The surrounding control flow decides what the model sees, which actions it may take, how failures change the next attempt and what survives when one context window ends. This atlas follows the architectures that made those decisions explicit, in date order.
Think of preparing a complex meal. A prompt chain is a recipe with fixed steps. A tool loop lets the cook inspect each result before choosing what to do next. A supervisor gives separate jobs to several cooks. A harness keeps the recipe, ingredients, timers, safety rules and progress notes organized, so another shift can continue the meal without guessing what happened.
Predict First
A coding agent must continue work after its current context window ends. What matters most for the next window?
Open an entry and read it in order: the limit it answered, what changed, then the mechanism in the diagram.
From prompt chains to durable harnesses
Search, or open a row to inspect the control flow, a concrete example and the primary source.
21 of 21 entries shown, in date order within each era.
Neural Mastery
2021–2022
Prompt programs
A single prompt was hard to inspect and could not reliably combine language with exact computation. Chains split work into visible stages, while modular routers sent each request to a suitable expert or tool.
2022–2023
Agents become loops
The program stopped deciding every step in advance. Models began choosing actions, reading results, revising plans and carrying useful state into the next attempt.
2023–2024
Teams, graphs and environments
One free-running loop was difficult to control. Search trees, role-based teams, explicit state graphs and purpose-built computer interfaces made control flow visible and constrained.
2024–2025
Production harnesses
Long tasks exposed a systems problem around the model. Production designs added routing, parallel workers, browser control, research loops, context curation and durable handoffs between separate context windows.
The Architecture Surrounds the Model
These entries did not replace one another. A production research agent can route the request, ask an orchestrator to create parallel searches, run each worker as a ReAct loop, compact the findings into a smaller context and store a cited report for the next session. The architecture is the set of choices that connects those parts and limits their behavior.
The diagrams show control flow, not model internals. When a system fails, identify which layer made the decision: the model, the tool interface, the router, the memory policy, the graph transition or the durable harness.