Blog

The Self-Improving Stack

Map the parts an agent team can change—prompts, skills, runtime, traces, evaluations, model training, and release rules—and the evidence required before promoting a candidate change.

For teams improving agents, the series helps choose the right layer, protect fresh test cases from the search process, and keep a human release decision at the end.

Start here: The Self-Improving Stack: How AI Agents Get Better. Then continue in the order below.

An editorial still life about how an agent system improves
Part 1

The Self-Improving Stack: How AI Agents Get Better

The self-improving stack is the set of agent layers that can change, from prompts and skills to runtime, traces, evaluation, model training, and governance.

An editorial still life about how an agent system improves
Part 2

Prompt Optimization for AI Agents: What It Can Change

Prompt optimization searches instructions, examples, schemas, tool descriptions, and rubrics for better agent behavior while the model and runtime stay fixed.

An editorial still life about how an agent system improves
Part 3

Skill Optimization for AI Agents: Durable Procedures

Skill optimization improves reusable agent procedures and their activation rules while keeping model weights fixed.

An editorial still life about how an agent system improves
Part 4

Optimization Theory for Agent Builders

Optimization theory for agent builders helps choose what to change, compare candidates fairly, and stop noisy gains from becoming production regressions.

An editorial still life about how an agent system improves
Part 5

Post-Training Agents: When to Change the Model

Post-training for agents moves improvement from prompts and skills into model weights or adapters, with harder data, release, and rollback decisions.

An editorial still life about how an agent system improves
Part 6

Agent Memory: Why Retrieval Is Not Learning

Agent memory becomes learning only when a scoped, evidence-backed write reaches the right future task and evaluation shows better behavior than no memory.

An editorial still life about how an agent system improves
Part 7

Evaluation Gates: Decide What Agent Changes Ship

Evaluation gates decide whether a change ships by comparing candidate and baseline on protected tasks, deterministic checks, cost, and trace evidence.

An editorial still life about how an agent system improves
Part 8

Agent Traces: The Evidence an Improving System Needs

Agent traces preserve calls, tool actions, observations, artifacts, costs, and outcomes so the team can diagnose failures, not tune scores alone.

An editorial still life about how an agent system improves
Part 9

Test-Time Compute for Agents: Beat Random at Equal Cost

Test-time compute adds samples, branches, retries, or checks; an execution shape matters only when it beats a simple baseline at equal cost.

An editorial still life about how an agent system improves
Part 10

Harness Evolution for AI Agents: When Prompts Plateau

Harness evolution changes the software around an agent when prompts and skills cannot add tools, isolation, checks, traces, or rollback.

An editorial still life about how an agent system improves
Part 11

Agent Runtime Topology: When Parallel Work Is Real

Agent runtime topology determines whether prompts produce parallel work, reviews, retries, cancellation, and evidence, or only describe them.

An editorial still life about how an agent system improves
Part 12

Multi-Agent Coordination: Roles Are Not Structure

Multi-agent coordination needs role contracts, state boundaries, budgets, checks, and traces; compare it with a single-agent baseline at equal cost.

An editorial still life about how an agent system improves
Part 13

Agent Governance: How to Govern Self-Improving Agents

Agent governance sets owners, authority limits, evidence requirements, approvals, and rollback paths before behavior changes reach users.