Research map · cutoff 29 July 2026

From agent loops to evolving systems

A tree of the expanding editable surface: output → context → memory → skills → workflow graph → harness → outer loop → source code → improvement operator → model–environment ecosystem.

594deduplicated papers, surveys, and primary blogs

240records dated 2026

11connected development branches

60core/frontier works closely reviewed

How to read the map

Solid edges show the primary lineage. Dashed edges show cross-branch dependence. Layout spacing is derived from direction, fan-out, and route congestion. Hover a node to inspect direct relations; double-click it to zoom and keep those relations in focus. Click the background to exit. Pinch or Ctrl/⌘ + wheel to zoom around the pointer.

Start here

The action loop is the trunk

2021–2023 · grounded interaction

Tool calls and environments create observable trajectories. Later branches decide what part of the system should learn from those trajectories.

Evidence status

peer-reviewed or accepted

preprint or technical report

official industry or practitioner source

survey / roadmap

conceptual branch or synthesis

Filter branches

Choose one or more · no selection shows all

2021–22

2023

2024

2025

2026

Concern-centered history

Five eras, defined by the question absorbing the field

This view intentionally suppresses most branch detail. Read left to right: each era identifies the dominant engineering question, the component people most wanted to change, and the bottleneck that pulled attention into the next era.

Dominant concernWhat most work was trying to unlock Editable surfaceWhat the system was newly allowed to change New bottleneckWhy the next era became necessary

2021–22External action interfaces

2023Persistent experience

2024Optimizable system graphs

2025Automating the agent designer

2026Governed evolving runtimes

ERA 01 2021–2022

Ground the model

Can a language model take useful actions and receive feedback grounded outside its own text?

Dominant concern

Action interfaces and observable trajectories

What most work cared about

browsing tool calls code execution human feedback

Editable surface

static response → action sequence

Representative work

WebGPT ReAct PAL Toolformer

New bottleneck A trajectory could improve the current answer, but successful experience disappeared after the run.

ERA 02 2023

Learn from experience

Can an agent improve across attempts without retraining the foundation model?

Dominant concern

Reflection, memory, reusable skills, and role specialization

What most work cared about

self-critique long-term memory skill libraries multi-agent roles

Editable surface

trajectory → persistent context

Representative work

Self-Refine Reflexion Generative Agents Voyager AutoGen

New bottleneck Experience was useful but unstructured; hand-designed memory, roles, and workflows did not scale across tasks.

ERA 03 2024

Optimize the system

Can prompts, modules, agent roles, and communication topology be searched instead of hand-authored?

Dominant concern

Composable agent programs and optimizable workflow graphs

What most work cared about

prompt compilers module search topology benchmarks & cost

Editable surface

memory/skill → workflow graph

Representative work

GPTSwarm DSPy TextGrad AgentSquare AFlow

New bottleneck The search space became combinatorial, while the evaluator still struggled to attribute which component caused a gain.

ERA 04 2025

Automate the designer

Can a system invent, edit, and preserve complete agent programs rather than tune a fixed template?

Dominant concern

Full agent code, open-ended archives, and long-running state

What most work cared about

agent-code search self-modification diverse archives scientific evaluators

Editable surface

workflow graph → source code + lineage

Representative work

ADAS Gödel Agent DGM AlphaEvolve Long-running harnesses

New bottleneck Self-editing exposed regressions, context resets, evaluator overfit, high search cost, and the need for rollback.

ERA 05 2026

Govern the evolving runtime

Can context, memory policy, workflow, harness, model, and environment improve together without losing control?

Dominant concern

Adaptive runtime systems with evidence-gated promotion

What most work cares about

active context trained memory updates harness evolution outer loops governance

Editable surface

agent code → model + harness + world + gate

Representative work

ACE Context Folding ReasoningBank Code as Harness HGM July survey

Current bottleneck Trustworthy credit assignment: prove that a change transfers, is current, is safe, and did not merely exploit its evaluator.

The eras overlap

These are attention regimes, not hard boundaries. ReAct remains inside 2026 harnesses; memory, workflow search, and code evolution continue in parallel.

The 2026 fork

The capability path expands what may evolve. The control path strengthens evidence, lineage, sealed tests, permissions, and rollback. Production systems need both.

The central change

Agent evolution moved outside model weights. Context, memory, tools, graphs, runtime code, and acceptance rules are now explicit mutable artifacts.

Why graphs recur

Graphs expose composable nodes, dependencies, lineage, and typed mutation surfaces. Workflow graphs and experience graphs solve different problems.

The governing constraint

Mutation is easy to generate. Crediting a real improvement without overfitting, reward hacking, or regressions is the hard part.