concept · concepts/Agent Autonomy Spectrum.md

Agent Autonomy Spectrum

status: active · created 2026-06-02 · updated 2026-11-08 · 4 sources · #agents #harnesses #autonomy #workflow #ux

The range of how hands-off you are with an agent — from turn-based "assisted AI" through semi-synchronous harness work to mega-autonomy — with NLW arguing the productive sweet spot is settling in the middle.

Harnesses

A harness is "the interface through which we interact with agents." NLW credits 2026's agent breakthrough to model advances (late 2025) + a greater focus on harnesses. The harness, more than raw model capability, determines where on the autonomy spectrum a workflow lands.

The spectrum

Turn-based (assisted AI) — prompt → wait → review → prompt. Too much latency; "you sit around waiting for its response."
Semi-synchronous (the sweet spot) — harnesses like Codex and Claude Code where you spin up and steer multiple agent threads, see their work, and interrupt anytime. Lower latency between guidance and execution. Compressed further by steering features and voice input. (The Human Sandwich lives here.)
Mega-autonomy — OpenClaw-style: agents on a dedicated Mac mini/studio, interfaced via Telegram, driven by "heartbeats" (recurring timed self-reminders to keep progressing toward goals). Found to "not actually accomplish the things people wanted" and to "absolutely burn tokens" (Token Shortage).

Evidence of the migration

Matt Schumer wiped his OpenClaw Mac mini, repurposing it as an "always-on dev box to use with Codex Mobile."
Nick Bauman (OpenAI) runs Codex across phone + MacBook + always-on Mac mini, starting/resuming threads from any device — "semi-synchronous," "the shape of it feels very real."
Jason Liu (OpenAI)'s "nine tips for getting the most out of Codex" all reduce, NLW argues, to parallel processing and semi-synchronicity rather than a turn-based paradigm (AIDB — Nine Tips for Getting the Most Out of Codex #to-read).
Manus — in Ethan Mollick's testing, this Claude-based agent showed "a tendency to extrapolate answers without asking," a concrete argument for keeping a human in a semi-synchronous loop rather than going fully autonomous (Making AI Work — Leadership, Lab, and Crowd).

Measured autonomy (science usage, 2026)

Anthropic — How Scientists Use Claude measures autonomy instead of arguing for it.

Its survey scale (none → assistive → shared → delegated → autonomous) is a finer-grained version of this spectrum.
In biology and biomedicine, the share of sessions run with high AI autonomy is 23.7% on Claude (chat), 64.1% on Claude Code, and over 70% on the specialized Claude Science.
Higher-capability subfields hand over more autonomous work (AI Use Tracks AI Capability).
Its acceleration scenarios find today's gains limited because use is mostly assistive or shared. Bigger multiples (8–16×) require moving toward full automation of non-physical steps (AI-Accelerated Science).

So the good setting on the dial depends on the task: verifiable, capability-rich work moves toward autonomy, and the rest stays semi-synchronous.

A multi-agent harness in practice

Anthropic — Investigating Vault Particles with Claude describes an internal multi-agent harness with access to internal compute, used to set up and run RELION (cryo-EM processing) end to end — "hundreds of GPUs in parallel." The autonomy setting is the semi-synchronous middle, stated plainly: Claude ran each step while the humans "advised on the analysis decisions." The one time it ran ahead on its own — skipping a standard processing step — it produced a map blended with its mirror image, and the correction was an explicit narrowing of autonomy: "Please let's do normal things unless they don't work." That is the Manus "extrapolate without asking" failure mode, in a lab, with a measurable cost.

The hardware footnote cuts against mega-autonomy economics from a different direction than Token Shortage: the compute-heavy step was the only part of the project a typical lab couldn't reproduce.

Connections

Verification Requires Domain Knowledge — how far the dial can safely turn depends on what the supervisor can check.
AI Use Tracks AI Capability — autonomy rises with task capability; harness specialization raises it further.
Codex, Claude Code — the harnesses anchoring the semi-synchronous middle.
OpenClaw — the mega-autonomy end that underdelivered.
Token Shortage — the economic argument against always-on autonomy.
The Human Sandwich — semi-synchrony is what makes the frame→collapse→judge loop tight.
Manus — a concrete autonomous agent whose "extrapolate without asking" failure mode argues for the semi-synchronous middle.
Communication and Control Are Intertwined — see also: steering/voice and heartbeats are feedback-loop control via communication.
Match Process Weight to Stage and Stakes seed — structural twin: a "match the dial to the context" heuristic, but the dial is process weight rather than agent autonomy.

Sources

The AI Daily Brief — Why Agents Still Need Humans — NLW synthesizing OpenClaw vs. Codex/Claude Code work patterns.
Making AI Work — Leadership, Lab, and Crowd — Mollick's Manus test (the agent failure mode).
Anthropic — How Scientists Use Claude — measured autonomy shares by interface; five-level survey scale.
Anthropic — Investigating Vault Particles with Claude — an internal multi-agent harness run semi-synchronously, and the cost of one unsupervised improvisation.

Linked from 19

Agent Memoryauto-ingested

Agent Autonomy Spectrum — long-horizon autonomy depends on memory that survives across sessions.

Agents as Employees vs Human-Agent Collaboration

Agent Autonomy Spectrum — Mode 1 sits toward higher autonomy; Mode 2 is explicitly semi-synchronous.

AI Use Tracks AI Capabilityauto-ingested

Agent Autonomy Spectrum — autonomy also rises with capability.

AIDB — Nine Tips for Getting the Most Out of Codexto read

NLW argues the nine tips "pretty much all come back to how to better parallel process and live in a state of semi-synchronicity with your agents instead of being stuck in some turn-based paradigm." Ingesting it would ground the semi-synchronous middle of th…

Anthropic — How Scientists Use Claudeauto-ingested

Agent Autonomy Spectrum — the survey's none → assistive → shared → delegated → autonomous scale, and autonomy rising with capability and with specialized tools.

Anthropic — Investigating Vault Particles with Claudeauto-ingested

Agent Autonomy Spectrum — an internal multi-agent harness run semi-synchronously.

Claude Code

Sits in the semi-synchronous middle of the Agent Autonomy Spectrum.

Codex

OpenAI's agentic coding/work harness — one of the tools anchoring the semi-synchronous middle of the Agent Autonomy Spectrum.

Communication and Control Are Intertwined

Agent Autonomy Spectrum — see also (weak resonance, not grounding): agent "steering," voice input, and OpenClaw "heartbeats" are control exercised through communication / feedback loops — a possible modern instance of this coupling.

Manus

Agent Autonomy Spectrum — a concrete autonomous agent; the "extrapolate without asking" failure mode is an argument for the semi-synchronous middle.

Match Process Weight to Stage and Stakes

Agent Autonomy Spectrum — structural twin: both are "match the dial to the context" heuristics.

OpenAI

Maker of Codex, cited as a human-agent collaboration harness; Nick Bauman and Jason Liu (both OpenAI) are sources for the semi-synchronous, multi-device Codex workflow (Agent Autonomy Spectrum, The AI Daily Brief — Why Agents Still Need Humans).

OpenClaw

Agent Autonomy Spectrum — OpenClaw marks the mega-autonomy end of the spectrum.

Personal vs Team Agents

"Every time an agent broke, the person it belonged to had to fix it." Even with a stable harness, agents need upkeep.

The AI Daily Brief

AIDB — Nine Tips for Getting the Most Out of Codex #to-read — semi-synchronous Codex workflow.

The AI Daily Brief — Why Agents Still Need Humans

The autonomy spectrum is settling toward the middle (Agent Autonomy Spectrum): early-2026 OpenClaw maximal autonomy (Mac minis, Telegram, "heartbeats") mostly burned tokens — relevant in the Token Shortage era — and people (Matt Schumer, Nick Bauman) are mi…

The Human Sandwich

Agent Autonomy Spectrum — the sandwich implies semi-synchronous working, not full autonomy.

The Infinite Backlog

Agent Autonomy Spectrum — OpenClaw-style "heartbeats" tried to chew through the backlog autonomously; many found it burned tokens without doing what they wanted.

Token Shortage

This is part of why work patterns are migrating toward more deliberate, semi-synchronous use.

Links to 17

Codex
Claude Code
The Human Sandwich
OpenClaw
Token Shortage
AIDB — Nine Tips for Getting the Most Out of Codexto read
Manus
Ethan Mollick
Making AI Work: Leadership, Lab, and Crowd
Anthropic — How Scientists Use Claudeauto-ingested
AI Use Tracks AI Capabilityauto-ingested
AI-Accelerated Scienceauto-ingested
Anthropic — Investigating Vault Particles with Claudeauto-ingested
Verification Requires Domain Knowledgeauto-ingested
Communication and Control Are Intertwined
Match Process Weight to Stage and Stakes
The AI Daily Brief — Why Agents Still Need Humans

In the log 3

2026-11-08 auto-ingest

Anthropic — Investigating Vault Particles with Claude (Anthropic Life Sciences, 2026-11-04; subscribed-feed capture) — unattended; 1 source + 1 concept (Verification Requires Domain Knowledge — you can only catch AI errors where you already have expertise), both tagged auto-ingested. Deliberately lean per T5 salience. The wiki's first worked case of AI-accelerated science: prokaryotic vault structure (phage L05, 96 copies, 9.5 MDa) in <4 months by a team outside its expertise; purification 3 days → 90 min, first model in 34 h; Claude's errors documented (2 weeks lost to a bad protocol; mirror-image cryo-EM map; "we told it to stop improvising"). Updated AI-Accelerated Science (case study + 1 contradiction: physical bench time treated as the cap, but a method substitution cut a physical step ~48× with no lab automation), AI Use Tracks AI Capability (verifiability has two halves — ground truth exists vs operator can read it), The Bottleneck Shift (Claude supplied method ideation + a coenzyme-A hypothesis; also argued against the method that worked), The Human Sandwich (the pattern in a wet lab; the bread must be load-bearing), Agent Autonomy Spectrum (internal multi-agent harness; the one unsupervised improvisation), AI Commoditizes Expertise (borrowed expertise and its bill), Critical Thinking Is Domain-Bound (first field evidence on its open question), Human-in-the-Loop by Exception + Falsification Is Asymmetric (see-also, no source-count bump), Anthropic (in-house biology lab), What's Scarce When Execution Is Cheap (narrows the scarce contribution from judgment toward verification), Overview, index. Method Reflection: T3 sharpened again — review only catches what the reviewer knows, which lands directly on T4's PR-review checkpoint; 4 candidate counterweights recorded. Seed Match Process Weight to Stage and Stakes: third promotion-candidate callout (verdict: analogy only — fourth near-miss; needs a direct source). 3 needs-your-call (1 contradiction, 1 promotion candidate, 1 method question). qmd search still broken (Node ABI) — searched with grep/index.
2026-11-08 auto-ingest

Anthropic — How Scientists Use Claude (Anthropic Research / Economics, 2026-11-05; subscribed-feed capture) — unattended; 1 source, 2 concepts opening the AI & science thread (AI Use Tracks AI Capability, AI-Accelerated Science), all tagged auto-ingested. Updated The Bottleneck Shift (science usage evidence + 1 contradiction: is the what-to-do bottleneck durable? 72% expect AI to match top researchers at ideation by 2030), Individual vs Organizational AI Gains (adoption lags capability; J-curve), Agent Autonomy Spectrum (measured autonomy 23.7% / 64.1% / >70%), Token Shortage (AI cost caps acceleration ~16×), AI and Employment — Replacement vs Expansion (task-level evidence), Human-in-the-Loop by Exception (see-also, not grounding), What's Scarce When Execution Is Cheap, Anthropic (Claude Science, Economics team, Cowork), Claude Code (science usage; Cowork is a separate surface), Dario Amodei ("Machines of Loving Grace"), Overview, index. Method Reflection: sharpened T3 (AI takes analysis + writing, humans keep ideation; that split may not last). Seed Match Process Weight to Stage and Stakes: second promotion-candidate callout (verdict: analogy only). 3 needs-your-call (1 contradiction, 1 promotion candidate, ops). qmd search still broken (Node ABI) — searched with grep.
2026-06-20 capture

seed "match process weight to stage and stakes" → seeds/Match Process Weight to Stage and Stakes.md; 0 to-read stubs (human flagged no person/work); back-linked from Agent Autonomy Spectrum (structural twin) + The Bottleneck Shift; opens new process/methodology domain (grounding candidates noted: Boehm & Turner, Cynefin, Cockburn — not yet flagged)