Why Agents Slow Down and Cost More? Achieving True Self‑Evolution by Subtraction

1. Context Density Over Length

Many practitioners assume that a larger context window automatically solves long‑term memory problems. The article shows that “capacity” does not equal “findability” or “usability”. Experiments (the “Lost in the Middle” test) reveal a U‑shaped performance curve: when crucial evidence lies in the middle of a long prompt, model accuracy drops sharply because attention concentrates on the beginning and end.

Consequently, the first design principle is density > length : keep only information directly relevant to the current task, environment state, and next decision. A high‑density, indexable context yields more stable decisions than a verbose, narrative‑style prompt.

Quantitatively, some agent frameworks keep 200 k–1 M tokens in system prompts and tool definitions, whereas GenericAgent caps the persistent context at ~30 k tokens. By scanning a webpage, discarding hidden or irrelevant DOM nodes, and retaining only essential elements, token usage can be reduced by about 90 % compared with feeding the full page to the model.

2. Minimal Yet Complete Tool Set

Tool redundancy is identified as a “hidden killer”. Adding a tool introduces a description, a choice, and a potential error path. Usage data from Claude Code shows that a single AgentTool accounts for 50.4 % of calls, while dozens of specialized tools together contribute only a tiny fraction.

The article advocates a small, orthogonal toolbox: five categories and nine atomic tools (e.g., file_read, file_patch, file_write, code_run, web_scan, web_execute_js, update_working_checkpoint, start_long_term_update, ask_user). Each tool has a single, well‑defined responsibility, making selection easier and security auditing more feasible.

Example workflow: to “read a paper and generate a public‑account article”, the agent first uses file_read to locate abstract, methods, and experiments, then code_run to extract figures, web_scan to verify public data, and finally file_write to compose the article. No monolithic “paper‑to‑article” tool is needed; the task is solved by composing a few atomic capabilities.

3. Hierarchical, Verifiable Memory

What many call “long‑term memory” is merely replaying past chat logs, which adds noise without value. The proposed memory architecture has five layers:

Meta‑rules that enforce system boundaries.

Ultra‑minimal index for fast navigation.

Global facts that store stable environment information.

Task‑level skills that capture reusable procedures.

Conversation archives for reflection and back‑tracking.

Higher layers contain short, stable pointers; lower layers hold detailed data loaded on demand. This pyramid ensures that adding new experiences does not inflate the always‑resident context.

4. Evolution Governed by Action and Verification

Self‑evolution is risky if unverified guesses become permanent knowledge. Four “iron laws” are enforced:

No action, no memory : only information verified through real tasks is written to long‑term memory.

Preserve verified data : compression may rephrase facts but must not discard verified evidence.

Avoid volatile state : timestamps or session IDs are never stored in the long‑term layer.

Minimal sufficient pointers : upper layers keep only the shortest identifiers needed to locate lower‑layer knowledge.

John Dewey’s quote is used to illustrate that learning comes from reflecting on experience, not merely from the experience itself. The pipeline transforms raw execution logs into SOPs (Standard Operating Procedures) that become reusable skills.

5. Measuring True Evolution

Evolution must be measurable: repeated execution of the same task should show higher success rates, lower latency, reduced token cost, and fewer human interventions. In benchmark tests, GenericAgent consumes only 27.7 % of the tokens used by Claude Code and 15.5 % of those used by OpenClaw , while achieving higher task completion rates.

Five consecutive runs of a task demonstrate that most agents start from scratch each time, whereas GenericAgent refines its process into an L3‑level SOP, continuously improving efficiency. The key metric is whether experience “crosses task boundaries” – i.e., whether the agent learns to solve new variations faster after earlier runs.

6. The Four‑Dimensional Value Function

To steer autonomous exploration, a weighted value function is proposed: 35 % real utility, 25 % capability breadth, 25 % depth, and 15 % innovation. Real utility dominates to keep evolution demand‑driven; breadth fills capability gaps; depth pushes existing skills toward “can do” → “do well”; innovation encourages cross‑domain composition.

Without demand‑driven goals, self‑evolution degenerates into self‑indulgent behavior.

Overall, the article concludes that higher‑level agents must practice restraint: they should manage context tightly, keep toolkits minimal yet composable, and let experience grow through validated actions, thereby achieving true self‑evolution that is both efficient and reliable.