The Self-Improving Stack

---title: 'The Self-Improving Stack'description: 'A series map for self-improving agent systems, from optimization theory and prompt search to runtime topology, traces, memory, and governance.'date: 2026-06-05tags: ['agents', 'evals', 'systems', 'self-improvement']draft: falsefigure:  src: '/images/software-3.svg'  alt: 'Software 1.0: source code. Software 2.0: learned neural network weights. Software 3.0: natural-language prompts.'  caption: 'Three representations of a program, after Andrej Karpathy. Agent systems can combine all three.'  source: 'https://www.youtube.com/watch?v=LCEmiRjPEtQ&t=85s'series: 'the-self-improving-stack'outline_trace_id: '2026-06-05T12-08-35-196Z-gpt-5.5'human_takeover: 'complete'authors:  - model: 'gpt-5.5'    role: 'outline'    date: 2026-06-05  - { model: 'gpt-5.5', role: 'draft', date: 2026-06-05 }  - { model: 'gpt-5.5', role: 'polish', date: 2026-06-05 }  - { model: 'gpt-5.5', role: 'review', date: 2026-06-05 }  - { model: 'gpt-5.5', role: 'publish', date: 2026-06-05 }  - { model: 'gpt-5.5', role: 'rewrite', date: 2026-06-05 }  - { model: 'gpt-5.5', role: 'polish', date: 2026-06-08 }  - { model: 'gpt-6-astra', role: 'diagram', date: 2026-10-02 }  - { model: 'gpt-6-luna', role: 'polish', date: 2026-10-02 }revisions:  - { date: 2026-10-02, model: 'gpt-6-luna', role: 'polish', note: 'Replaced prose code blocks with lists, equations, and compact flows. Audited recovery: selected public tool inputs from the October 2 editing session, not the complete rollout. Parent integration corrected MDX, display math, and responsive layout; full native records remain private.', commit: '99791a3a1484aedf2f2cda6d56b3421b5c354f0a', trace_id: '2026-10-02T23-45-17-233Z-gpt-6-luna-the-self-improving-stack-polish' }  - { date: 2026-10-02, model: 'gpt-6-luna', role: 'polish', note: 'Rendered existing equations with KaTeX. Audited recovery from the Luna editing session: selected public messages and tool-input previews, not the complete rollout. Parent integration review corrected prime notation.', commit: 'a68bc06ef65efe0e41ede1d33205cda3a494f38e', trace_id: '2026-10-02T22-42-45-776Z-gpt-6-luna-the-self-improving-stack-polish' }  - { date: 2026-10-02, model: 'gpt-6-astra', role: 'diagram', note: 'Added a Software 3.0 figure, reused in the article and link preview; prose unchanged. This record is a selected session excerpt; the full source remains private.', commit: '552acee10e8bbb22e7761d0807565eaac2c8d5a2', trace_id: '2026-10-02T21-57-48-883Z-gpt-6-astra-the-self-improving-stack-diagram' }  - { date: 2026-06-08, model: 'gpt-5.5', role: 'polish', note: 'clarified that self-improvement targets the user-task distribution, tightened the loop equation, and tied evidence to task outcomes', commit: 'd0bb565e0643eb9389876935b5191b5482c9db38', trace_id: '2026-06-08T10-10-44-256Z-gpt-5.5-the-self-improving-stack-polish' }  - { date: 2026-06-05, model: 'gpt-5.5', role: 'polish', note: 'let''s track a section for each of these map items, and eventually a full article too, but i want a directory we can use to checkpoint our kn · 37 asst turns · 23 tool calls', commit: 'fb31e1c764d9711386702764aaf1c2c5cf9886aa', trace_id: '2026-06-05T12-35-48-868Z-gpt-5.5-the-self-improving-stack-polish' }  - { date: 2026-06-05, model: 'gpt-5.5', role: 'rewrite', note: '60/40 voice rewrite: grounded openings in concrete agent-work failures, removed scaffold headings, added falsification pressure, and tightened paragraph rhythm while preserving source trails.', commit: 'd5bba9f0c633e5d2794e9b8e062ab48b15bbd1f5', trace_id: '2026-06-05T12-35-48-868Z-gpt-5.5-the-self-improving-stack-rewrite' }  - { date: 2026-06-05, model: 'gpt-5.5', role: 'publish', note: 'Published the self-improving stack series at Drew''s request, marking human takeover complete and flipping the post live.', trace_id: '2026-06-05T12-08-35-196Z-gpt-5.5-the-self-improving-stack-publish' }  - { date: 2026-06-05, model: 'gpt-5.5', role: 'review', note: 'Reviewed and dated the source trail, removed remaining temporal language, and marked the source-freshness checkpoint complete.', commit: 'b8fd3dbe812dd9ddd73865ae65fcc0d381b59d69', trace_id: '2026-06-05T12-08-35-196Z-gpt-5.5-the-self-improving-stack-review' }  - { date: 2026-06-05, model: 'gpt-5.5', role: 'polish', note: 'Polished the umbrella article with a layer-confusion diagnostic, tightened promotion-gate phrasing, and verified role-scoped trace capture for separate draft and polish provenance.', commit: 'b8fd3dbe812dd9ddd73865ae65fcc0d381b59d69', trace_id: '2026-06-05T12-08-35-196Z-gpt-5.5-the-self-improving-stack-polish' }  - { date: 2026-06-05, model: 'gpt-5.5', role: 'draft', note: 'Drafted the umbrella series article with the closed-loop formalism, layer table, practical test, and full series map; replaced outline handoff prose and synced the research overview/status map.', commit: 'b8fd3dbe812dd9ddd73865ae65fcc0d381b59d69', trace_id: '2026-06-05T12-08-35-196Z-gpt-5.5-the-self-improving-stack-draft' }  - date: 2026-06-05    model: 'gpt-5.5'    role: 'outline'    note: 'Research planning pass from a traced session.'    trace_id: '2026-06-05T12-08-35-196Z-gpt-5.5'supporting_trace_ids:  - '2026-06-05T12-08-35-196Z-gpt-5.5'---import CandidateLoop from '../../components/CandidateLoop.astro';import Steps from '../../components/Steps.astro';I can tell a coding agent to parallelize work, and it will often agree with me while still doing one thing at a time.That failure looks like a prompting problem until you inspect the trace. The sentence "fan out independent subtasks" changed the model's intention, but it did not create a worker pool, a scheduler, a merge rule, a verifier, or a budget policy. The prompt moved. The action space did not.That is the category error hiding inside a lot of talk about self-improving agents. We say:The system optimizes itself.as if there were one surface called "the system."There is not.There are prompts, skills, tools, traces, memory stores, evaluators, runtime graphs, harnesses, model weights, and release gates. Each one can be optimized. Each one needs a different kind of evidence. Each one can fail in a different way.Self-improvement only has content after you name the task class. The target is better execution of the work the user is trying to get done: the code change, research answer, design review, deployment, diagnosis, or decision that caused the agent to be invoked in the first place. A system that improves a judge score while making that work slower, less faithful to intent, or harder to audit has optimized a proxy, not the task.The useful questions are more concrete:- Which user task distribution is being improved?- What is allowed to change?- What evidence shows improvement on that task?- How are candidates generated?- What gate decides promotion?- What can go wrong when that layer changes?Those six questions are the self-improving stack.## The Loop Behind The WordA self-improving agent system has a closed loop:

# 00 Series OverviewPost: `src/content/posts/the-self-improving-stack.mdx`Status: publishedLast updated: 2026-06-06Supporting trace: `2026-06-05T12-08-35-196Z-gpt-5.5`Source freshness checked: 2026-06-06## Core QuestionWhat does it mean for an agent system to improve itself when model weights aremostly frozen and the mutable state lives in prompts, skills, traces, topology,tools, memory, evals, and harness code?## Claim To TestSelf-improvement is not a property of a model. It is a property of a closed loop:run, observe, diagnose, propose, validate, promote, remember, and govern. Thesystem is only as real as its feedback signal, trace integrity, promotion gate,and safety case.Formal update:```texts_{t+1} =  Promote(s_t, c_t)  if Gate(Eval(Run(c_t), Run(s_t)), policy) passes  else s_t```where:```texts_t = current system statec_t = candidate stateRun(.) = full trajectory under scenariosEval(.) = measured evidenceGate(.) = promotion rule under policy```## Five QuestionsFor every layer:- What is allowed to change?- What evidence says it improved?- How are candidates generated?- What gate decides promotion?- What failure mode does this layer create?## Series Spine| # | Layer | Mutable surface | Gate ||---|---|---|---|| 01 | Optimization theory | candidate state and objective | statistical and structural validity || 02 | Prompt optimization | instructions, examples, rubrics, LM program text | held-out prompt eval || 03 | Skill optimization | reusable procedures and action policies | transfer and invocation tests || 04 | Runtime topology | drivers, fanout, reviewers, selectors, turn budgets | trace integrity plus budget gate || 05 | Multi-agent coordination | roles, contracts, supervisors, workers | role isolation and selector audit || 06 | Test-time compute | samples, branches, retries, verifier calls | Pareto dominance under equal compute || 07 | Evaluation gates | scorecards, judges, baselines, release criteria | fail-closed promotion || 08 | Trace systems | spans, artifacts, raw calls, replay records | capture integrity and replayability || 09 | Harness evolution | source code around the agent | release gate outside the mutation surface || 10 | Post-training | model weights or adapters | model release and data-governance gate || 11 | Memory and knowledge | persistent state across episodes | source, freshness, scope, poisoning checks || 12 | Governance | authority, risk controls, release policy | accountable approval and rollback |## Converged Series Rules- Same outer skeleton does not mean same optimizer. The decisive difference is  the mutable surface.- Prompt optimization can improve wording inside a fixed runtime. It cannot add  action surfaces the runtime does not expose.- Skills are procedural memory, but they require activation and transfer gates.- Multi-agent systems are topology plus role contracts, not only personas.- Test-time compute gains need compute-matched baselines.- The gate is part of the optimizer because it defines what persists.- Traces are the learning data. Scores are lossy projections.- Harness evolution expands the reachable set but must not own its own release  gate.

# Self-Improving Agent SystemsThis directory is the checkpoint corpus for the blog series `the-self-improving-stack`.It is not publication prose. It is where research notes, claims, source trails,open questions, and article checkpoints live before they are turned into posts.Supporting trace: `2026-06-05T12-08-35-196Z-gpt-5.5`Source freshness checked: 2026-06-06## Core ClaimSelf-improving agent systems are not one technique. They are a stack of mutablesurfaces, feedback signals, search operators, and promotion gates. Confusioncomes from mixing layers: prompt optimizers mutate text, skill optimizers mutateprocedure, runtimes expose topology, eval systems decide promotion, meta-harnessesmutate architecture, and frontier tuning can change model/runtime behavior.The converged loop is:```textrunobservediagnoseproposevalidatepromoteremembergovern```Every article names the mutable surface, feedback signal, search operator,promotion gate, and failure mode for its layer.## How To Use This DirectoryFor each topic, keep the same questions current:- What is the mutable surface?- What feedback signal is trusted?- What search operator explores candidates?- What promotion gate prevents regression?- What failure mode does this layer create?- Which blog post currently owns the public-facing argument?## Series Map| # | Topic | Checkpoint | Draft post | Status ||---|---|---|---|---|| 00 | Series overview | `00-series-overview.md` | `the-self-improving-stack` | published || 01 | Optimization theory | `01-optimization-theory.md` | `self-improving-stack-optimization-theory` | published || 02 | Prompt and LM-program optimization | `02-prompt-lm-program-optimization.md` | `self-improving-stack-prompt-optimization` | published || 03 | Skill optimization | `03-skill-optimization.md` | `self-improving-stack-skill-optimization` | published || 04 | Agent runtime topology | `04-agent-runtime-topology.md` | `self-improving-stack-agent-runtime-topology` | published || 05 | Multi-agent coordination | `05-multi-agent-coordination.md` | `self-improving-stack-multi-agent-coordination` | published || 06 | Test-time compute | `06-test-time-compute.md` | `self-improving-stack-test-time-compute` | published || 07 | Evaluation and gates | `07-evaluation-and-gates.md` | `self-improving-stack-evaluation-gates` | published || 08 | Trace systems | `08-trace-systems.md` | `self-improving-stack-trace-systems` | published || 09 | Code and harness evolution | `09-code-harness-evolution.md` | `self-improving-stack-harness-evolution` | published || 10 | Model training and post-training | `10-model-training-post-training.md` | `self-improving-stack-post-training` | published || 11 | Memory and knowledge flywheels | `11-memory-knowledge-flywheels.md` | `self-improving-stack-memory-flywheels` | published || 12 | Safety, security, governance | `12-safety-security-governance.md` | `self-improving-stack-governance` | published |## Publication GatePublication order is numeric, starting with `the-self-improving-stack` as theseries map. Drew accepted the series for publication, so every post is now`draft: false` with `human_takeover: 'complete'`. The release gate was:```textpublish(post) iff  draft is ready  and source freshness is dated  and provenance rows exist  and anti-pattern scan is clean  and human_takeover != 'pending'```The current state satisfies the full release gate for the series.## Shared Vocabulary

#!/usr/bin/env node/** * trace-capture: harness-agnostic session capture for blog revisions. * * Usage: *   pnpm tsx tools/trace-capture.ts capture \ *     [--harness=claude-code|codex|manual] \ *     [--post=<slug>] \ *     [--role=outline|draft|rewrite|polish|diagram|review|publish|research] \ *     [--session=<session-id>] \ *     [--marker="<token>"] \ *     [--note=<one-line>] \ *     [--commit=<sha>] \ *     [--input=<path>]            # manual harness only *     [--kind=post|series-outline|supporting-research] *     [--attach=supporting|revision|none] *     [--latest]                  # choose latest session without requiring post file touch * *   pnpm tsx tools/trace-capture.ts capture --auto *     # detects from the latest git commit: finds changed posts, matches a *     # recent session via ~/.claude/projects or ~/.codex/sessions, writes a *     # trace per changed post, appends to frontmatter. * *   pnpm tsx tools/trace-capture.ts list *     # list existing traces grouped by post. * *   pnpm tsx tools/trace-capture.ts show <trace_id> *     # dump a trace as JSON. */import { execSync } from 'node:child_process'import { mkdir, readdir, readFile, writeFile } from 'node:fs/promises'import { join } from 'node:path'import ClaudeCodeHarness from './harness/claude-code.js'import CodexHarness from './harness/codex.js'import ManualHarness from './harness/manual.js'import { dedupeAdjacentTurns, type TraceFile, type TraceHarness, type Turn } from './harness/types.js'const ROOT = process.cwd()const POSTS_DIR = join(ROOT, 'src/content/posts')const TRACES_DIR = join(ROOT, 'traces')type Args = Record<string, string | boolean>function parseArgs(argv: string[]): { cmd: string; pos: string[]; flags: Args } {  const [cmd, ...rest] = argv  const flags: Args = {}  const pos: string[] = []  for (const a of rest) {    if (a.startsWith('--')) {      const eq = a.indexOf('=')      if (eq >= 0) flags[a.slice(2, eq)] = a.slice(eq + 1)      else flags[a.slice(2)] = true    } else pos.push(a)  }  return { cmd: cmd ?? 'capture', pos, flags }}function git(cmd: string): string {  try {    return execSync(`git ${cmd}`, { cwd: ROOT, stdio: ['ignore', 'pipe', 'ignore'] }).toString().trim()  } catch {    return ''  }}function headCommit(): string | null {  const sha = git('rev-parse HEAD')  return sha || null}function changedPostsAtHead(): string[] {  const out = git('show --no-renames --name-only --format="" HEAD')  return out    .split('\n')    .map((l) => l.trim())    .filter((l) => l.startsWith('src/content/posts/') && l.endsWith('.mdx'))    .map((l) => l.replace('src/content/posts/', '').replace(/\.mdx$/, ''))}