← The Independent 100: a study

dex, Kyle Mistele and HumanLayer (12-Factor Agents, ACE-FCA, skills, shannon, fold)

Deep dive, written 2026-09-25 by an AI agent session; public edition 2026-09-25 Covers dex (#87), Kyle Mistele (#97)

Bracketed tags are evidence labels: [Verified] means we read the page, file or API response ourselves, [Reported] means a source says so and we did not check the underlying thing, and [Inference] is our own reasoning. CI is continuous integration: automated checks that run on every push. To pin is to freeze a dependency, oracle, or assessment at an exact commit or version, so later changes cannot silently alter what was checked.

Credit

The Independent 100 list was curated by dan (@irl_danB) and published on 2026-09-24 at independent.prose.md (announcement). The ranks and section names are his, and number 100 is his open wild-card slot. This study is ours and is not endorsed by him or by anyone on the list.

Summary

How HumanLayer's 12-Factor Agents, ACE-FCA context compaction, control-loop skills, shannon and fold handle context, compaction and agent loops, and what they suggest for agent harnesses.

Sources

How to read this

Written 2026-09-25 by an AI agent session. [Verified] means we read the file at the pin or ran the check. [Reported] means the author or a third party says so and we did not check it. [Inference] is our reasoning.

What they built

12-Factor Agents (dex, 2025)

A doctrine repository of twelve factor files plus an appendix ("factor 13, pre-fetch context"); two generations of the files (factor-3-… and factor-03-…) coexist and differ. The code is a scaffolder and a set of workshops; the last commit was 2025-09-21. [Verified] The factors that bear on agent harnesses [Verified]:

  • Factor 3. "Everything is context engineering": a custom event format in place of the chat-message list, and resolved errors hidden (factor-03).
  • Factor 5. Infer execution state from the event log (factor-05).
  • Factor 9. Put the error into context and retry, but "limit to ~3 attempts of a single tool", then escalate (factor-09 line 31).
  • Factor 10. Keep agents to "3-10, maybe 20 steps max" (factor-10 line 9).

ACE-FCA and the essays (dex, 2025-2026)

The ACE repository holds four essays and no code. From ace-fca.md [Verified]:

  • The rule is "frequent intentional compaction", keeping utilization "in the 40%-60% range (depends on complexity of the problem)" (line 60, line 218).
  • The window it assumed comes from a Geoffrey Huntley quote: "approximately 170k of context window" (line 171).
  • The workflow is research, plan, implement; subagents are about context control, not role-play (line 191).
  • Humans review research and plans, because a bad line of research can produce thousands of bad lines of code (lines 318-333); "I *can* read 200 lines of a well-written implementation plan" (line 357).
  • Outcome claims we did not check [Reported]: "35k LOC … in about 7 hours" on BAML, "$12k on opus per month" for a team of three, and a failed parquet-java attempt.

wsff.md, "Why Software Factories Fail", argues "the lights off factory does not work" (line 205) [Verified]. The essay says "In July 2025 we went full lights-off" and that by November a full rewrite looked easier (line 226) [Reported]. Its explanation: training puts "no penalty for eroding codebase maintainability" (lines 347-349). Its proposed fix is four human steps before code: product review, system architecture, program design (call-stack trees, file-tree diffs, signatures) and vertical slices (lines 411-587). It closes with "Read the dang code" (line 618).

Two SlopCodeBench write-ups benchmark models over evolving checkpoints, each with a fresh context. On the write-up's 17-checkpoint subset Opus 5 reached a 24% strict pass (line 33). [Reported]

27 commands and 6 agents in the repository's Claude Code directory. [Verified]

  • create_handoff.md is a handoff template (tasks and status, critical references, recent changes as file:line, learnings, artifacts, action items) that prefers path:line references over code blocks (lines 45-64).
  • resume_handoff.md marks each recent change "Verified present/Missing/Modified" (lines 94-96) under the rule "Never assume handoff state matches current state … Verify all file references still exist" (lines 166-170), and names four scenarios: clean, diverged, incomplete, stale.
  • implement_plan.md pauses after each phase for manual verification (lines 50-63); create_plan.md splits success criteria into automated and manual (lines 345-371).

humanlayer/skills (MIT)

design-control-loop (SKILL.md) interviews the user, then builds a scheduled loop [Verified]:

  • set point, sensor, controller, actuator and disturbances, with a memory file loaded on every run;
  • "Make each component runnable locally and standalone before wiring it into CI" (line 28);
  • flow control: "Recommended default: one open PR per loop" (line 140), implemented with gh pr list --label (workflow-template.yml lines 55-63);
  • a dampener: a check that compares the sensor's output against a baseline and surfaces, or eventually blocks, newly introduced deviations while the loop chips away at the problem (line 75); advisory by default in the worked example (example-control-loop.md line 39).

improve-claude-md wraps CLAUDE.md sections in <important if="…"> blocks, on the premise that a Claude Code system-reminder calls the context "may or may not be relevant". [Verified: text read; the premise is not verified]

shannon (dex, MIT)

A Bun CLI and SDK that makes interactive Claude Code look like claude -p. [Verified]

  • It starts claude in tmux, waits for a prompt glyph in capture-pane output, and pastes the prompt with set-buffer, paste-buffer and send-keys C-m (index.ts lines 1028-1045).
  • A turn is done when an assistant row is followed by a system/turn_duration row (lines 694-718). Timeouts are 20 s to start and 180 s per turn (lines 78-80).
  • Parity with claude -p is marked "Partial" (GOAL_PROGRESS.md lines 70-71). The repository does not say why it avoids claude -p.

fold (Kyle Mistele, MIT)

A Bun monorepo on an Effect v4 release candidate: 409 TypeScript files across 13 packages [Verified]. On 2026-09-25 the GitHub API listed one contributor, the account K-Mistele (profile name Kyle Mistele, company HumanLayer: github.com/K-Mistele), with 212 commits, and 75 stars (github.com/humanlayer/fold) [Verified, 2026-09-25].

  • State is an append-only log of tagged entries (Schemas.ts lines 246-628), with read models as projections: 12FA factors 5 and 12 in code. [Verified]
  • Auto-compaction fires at window − min(32k, window/4) − min(16k, window/8) and keeps a 20k-token tail (CompactionEngine.ts lines 60-68). A stale-usage guard allows no second compaction until a fresh response reports usage (lines 83-103), and a context-overflow error triggers one compact-and-retry (lines 318-344). Agents may override the threshold. The summary prompts are "ported verbatim from pi" (CompactionPrompts.ts). [Verified]
  • After compaction the model gets the path of the full pre-compaction log, told to "Treat it as last-resort archival memory", with a narrow search recipe scoped to its own agent id (CompactionArchiveAccess.ts lines 36-58). [Verified]
  • Doom-loop stop: a run ends gracefully after N identical consecutive tool-call batches (StopConditions.ts); the default is 3 (Launch.ts lines 202-203). [Verified]
  • Claude models get write/edit tools and GPT/Codex gets apply_patch; an --rpi flag adds six specialist subagents ported from "riptide-rpi" prompts (Rpi.ts). [Verified]

Where the 40-60% figure went later

The March 2026 HumanLayer blog post (byline "Kyle"; we take it to be Kyle Mistele [Inference]) calls the low band the "smart zone" and the high band the "dumb zone", arguing an extended-context model is "the same model with some clever math", so a bigger window "just makes the haystack bigger". [Reported] The third-party QRSPI summary reports "keep context under 40%, start fresh at 60%" as practitioner consensus, and says QRSPI hides the ticket from research so research records facts, not opinions. [Reported, secondhand]

Lessons for agent harnesses

Comparisons below are [Inference] unless labelled.

Context target. They keep utilization at 40-60%, stated for windows of about 170k-200k, and the "Skill Issue" post argues bigger windows do not help. By default OMP auto-compacts at window minus max(15%, 16,384 tokens) (compaction.ts lines 333-335, lines 388-411) [Verified]. Their number is argued, not measured, for 1M windows.

Context sensor. fold reads context size from API-reported usage rather than estimating it from terminal scrollback. A harness that records the true number every turn gives a better sensor than any scrollback estimate.

When compaction happens. They compact intentionally at phase boundaries, though fold's automatic threshold still sits near the limit (about 95% of 1M, 76% of 200k). Automatic compaction near the limit is a safety net in both; the difference is workflow.

After compaction. fold tells the model where the raw log is. OMP offers a raw-transcript address (history://current/full) only in an experimental mode that is off by default (experimental-context-rollover.md; settings-schema.ts line 2528) [Verified]. Telling the model where the raw log is, as fold does, is cheap, bounded and better than guessing [Inference].

Post-compaction recovery. They have explicit /create_handoff and /resume_handoff commands. OMP exposes compaction events to extensions (types.ts lines 1252-1257) [Verified], so a harness can run the same kind of recovery step when compaction happens.

Verify before resuming. They mark every change a handoff names as present, missing or modified. That check keeps a resumed session from trusting a stale handoff.

Human review. They put human review at research and plan; their essay reports that lights-off failed [Reported]. Their evidence is one team's reported experience. Whether agent review can stand in for human review at the plan stage is open, and it is a policy choice for a human, not an agent.

Plan size. Theirs is about 200 lines, short enough for a human; QRSPI reportedly says 1,000-line plans hide surprises [Reported]. The right size depends on the reader: a plan a human reviews differs from one only agents read.

Loop design. They have a dampener, one open PR per loop, a memory file and local-first components. The dampener guards against regressions while a loop works and the one-PR bound limits work in progress; stall detection alone covers neither.

Doom loops. fold stops after 3 identical batches; OMP steers after 5 identical calls (settings-schema.ts lines 1383-1415) [Verified]. Where identical calls can be legitimate polling, steering suits better than stopping.

Repeated errors. Factor 9 says about 3 attempts, then escalate. A consecutive-error count is a useful signal even where no cap interrupts the agent.

Detecting turn end. shannon reads the transcript's turn_duration row; comparing pane captures and handling spinners is the text-heuristic alternative. A typed signal on disk beats text heuristics.

Agent size. 12FA says 3-10, maybe 20 steps. The unit matters: a small focused agent can be one task inside a long-lived session, so step caps and session lengths are not directly comparable.

Reuse terms

  1. 12-Factor Agents prose, images and snippets. CC BY-SA 4.0 §3(b) requires adaptations to carry BY-SA, and §2(a)(5)(B) forbids added downstream restrictions (legal code). [Inference: our reading, not legal advice] Paraphrasing and linking the factor avoids this; code under packages/ is Apache-2.0 and fine with a notice.
  2. ACE and wsff prose or images. No license file, so link only.

Attribution we owe

  • dex (github.com/dexhorthy; the ACE essays were committed from that account, per the commit history): "frequent intentional compaction", the 40-60% utilization target, research-plan-implement with human review at research and plan, and "subagents are about context control" (ACE). Cite the essays wherever these ideas are used.
  • Geoffrey Huntley (named in ACE as the source): the "approximately 170k of context window" quote and the Ralph loop that ACE cites. Covered in our separate Huntley deep dive.
  • HumanLayer's RPI commands (Apache-2.0): the handoff and resume structure and the verify-before-acting rule. If text is adapted, keep the Apache notice and state the changes (Apache-2.0 §4(b)-(c)).
  • Kyle Mistele, fold (MIT; sole contributor per the GitHub API): the post-compaction archive pointer, the stale-usage guard and the doom-loop stop. fold's compaction prompts are "ported verbatim from pi", so the chain runs back to Pi (MIT, copyright Mario Zechner), which OMP also descends from. Also the "smart zone / dumb zone" framing from the "Skill Issue" post.
  • humanlayer/skills (MIT): the control-loop vocabulary applied to agent loops (dampener as a baseline regression gate, one open PR per loop as flow control).
  • shannon (MIT, dex): detecting turn end from the transcript's turn_duration row.
  • 12-Factor Agents (CC BY-SA 4.0 content): link-level citation of factors 3, 9 and 10.
  • SlopCodeBench, Gabriel Orlanski and co-authors per the arXiv author list (arxiv.org/abs/2603.24755; repository MIT, problems Apache-2.0).
  • Chroma's context-rot report (Kelly Hong, Anton Troynikov and Jeff Huber, research.trychroma.com/context-rot), which the "Skill Issue" post cites as the empirical basis for the dumb zone.

What we could not verify

  • Outcome claims in ACE and wsff: BAML "35k LOC in 7 hours" and its merged PRs, "$12k on opus per month", the lights-off failure and rewrite, and the Faros AI figures. We did not check the PRs or the report.
  • The QRSPI content (eight stages, hiding the ticket during research, "under 40%, fresh at 60%"). We read only a third-party summary, not the talk.
  • Whether 40-60% fill harms quality on 1M-token windows. A comparative study is proposed and not yet run.
  • Running the subject code. We did not run shannon, fold, design-control-loop or any 12FA code, or their tests.
  • Provider terms for driving an interactive Claude Code session from tmux, as shannon and ntm both do.
  • The improve-claude-md premise that Claude Code under-weights CLAUDE.md sections because of a system-reminder.

How to read this

A deep dive compares one person’s public work with how we work and says what we would adopt. Its proposals are our plans, not recommendations to anyone else, and its judgements of fit are one reviewer’s. Source: study/independent-100/deep/dex-humanlayer.md.

Request a correction

Is something about this deep dive wrong or out of date, or would you rather not be covered? Open a correction (a GitHub issue form with this page’s address filled in). A request to remove a record is honoured without argument; accepted corrections are recorded in the changelog with the date.

Help fix it

Each opens a GitHub issue form with this page’s address filled in. You need a GitHub account.

Made by Joshua Nowak (ZestStream) with AI coding agents. Not affiliated with or funded by Jeffrey Emanuel; Joshua uses his public tools and is a paying subscriber to jeffreys-skills.md.

Source on GitHub MIT licensed