24 Oct 2026
Baseline
The newest METR point in v1.1 is dated 7 Apr 2026. Self-replication remains text-only (there is no maintained series).
← Research themes · Last updated 24 Oct 2026
Agent swarms and the "DNA" of a resilient AI agent
A conceptual map of which traits make agent populations robust, borrowed from biology, and what that implies for risk.
Thesis being trackedThe traits that make an AI agent or swarm resilient (persistent goals, memory, tools, cooperation, redundancy and replication with selection) are the same ones biology uses. The same traits that add resilience also make the system harder to control, so they need tracking as capabilities rise.
Conceptual and high-level only. This page describes traits and research findings. It contains no instructions for building self-replicating, evasive or autonomous agents.
The trait table
9 rows
| Trait | Biological analogue | Resilience benefit | Risk |
|---|---|---|---|
| Goals / values | Genome-encoded drives; kin selection | Coherent behaviour across contexts | Goal misgeneralisation; instrumental drives such as self-preservation (Carlsmith) |
| Memory | Neural memory; epigenetics | Learns from experience and keeps context | Accumulates errors or hidden state that is hard to audit |
| Tools | Tool use in primates and corvids | Extends capability without retraining (Voyager) | Larger action surface and more real-world impact |
| Self-modification | Development; plasticity | Adapts to new tasks (Darwin Gödel Machine) | Drift from the original specification; reward hacking (Lehman et al.) |
| Resource acquisition | Foraging; metabolism | Keeps running under scarcity | Competes with humans for compute, money and energy |
| Cooperation | Eusocial insects; superorganisms | Division of labour and collective problem-solving (swarm intelligence) | Collusion and emergent multi-agent failures (Hammond et al.) |
| Redundancy | Duplicate organs; distributed colonies | Survives failure of individual nodes | Harder to shut down |
| Replication with variation and selection | Darwinian evolution; r/K strategies | Open-ended improvement | Selection favours selfish traits (Hendrycks); loss of control. The red line in self-replication evaluations. |
| Immune-like monitoring | Adaptive immune system (self/non-self) | Detects and corrects faulty or compromised agents (artificial immune systems) | False positives; monitors can be gamed |
What to watch
- Agent frameworks adding persistent memory and self-improvement loops.
- Research on safeguards for multi-agent systems.
Reading
- Multi-Agent Risks from Advanced AI— Hammond et al. (Cooperative AI Foundation), 2025-02A taxonomy of multi-agent failure modes: miscoordination, conflict, collusion.
- Generative Agents: Interactive Simulacra of Human Behavior— Park et al., 2023The memory and reflection architecture behind many agents.
- Voyager: An Open-Ended Embodied Agent with Large Language Models— Wang et al., 2023Skill libraries as accumulated "tools".
- Darwin Gödel Machine— Zhang et al., 2025-05Self-improving agents using open-ended evolution.
- AlphaEvolve— Google DeepMind, 2025-05Evolutionary search with LLMs applied to real problems.
Biology analogies
Evolution
Variation plus selection plus inheritance produces adaptation without a designer. Digital-evolution experiments repeatedly found unexpected and sometimes specification-gaming solutions (Lehman et al.). Hendrycks argues competitive pressure selects for selfish AI traits (2023).
r/K selection
r-strategists produce many cheap offspring; K-strategists invest in few robust ones (Pianka 1970). The analogy: many disposable agent instances versus a few heavily resourced, persistent agents.
Immune systems
Distributed detection of "non-self" with memory of past threats. It inspires anomaly detection and monitor agents (artificial immune systems).
Superorganisms / ants
Simple agents following local rules give robust collective behaviour, the basis of ant-colony optimisation (overview). There is no central point of failure, which also means no central off-switch.
What to watch
- Papers applying evolutionary or selection framing to AI agent ecosystems.
Reading
- On r- and K-Selection— Pianka, The American Naturalist, 1970The classic r/K paper.
- The Surprising Creativity of Digital Evolution— Lehman et al., 2018Anecdotes of evolution gaming its objectives.
- Natural Selection Favors AIs over Humans— Dan Hendrycks, 2023-03The Darwinian argument for AI risk.
- Swarm intelligence— Wikipedia, 2026Background on collective behaviour.
Tracking indicators
METR time horizon is the length of software tasks (measured in human expert time) that frontier models complete with 50% (p50) or 80% (p80) reliability. METR reported it doubling roughly every seven months since 2019 (Kwa et al. 2025).
METR 50% time horizon of frontier AI agents (state-of-the-art models)
minutes of human-expert task time
Length of software/research tasks (measured in how long they take skilled humans) that a model completes with 50% (or 80%) reliability. Point estimates only; METR publishes wide confidence intervals. Each point is a model METR flags as state-of-the-art at release.
Source: METR, Time Horizon v1.1 benchmark results [2] [3]
· Data as of 7 Apr 2026 · fetched 24 Oct 2026
METR results by model
26 rows
| Model | Release | p50 horizon (min) | 95% CI low | 95% CI high | p80 horizon (min) | Frontier at release |
|---|---|---|---|---|---|---|
| gpt2 | 14 Feb 2019 | 0.1 | 0.0 | 0.1 | 0.0 | Yes |
| davinci_002 | 28 May 2020 | 0.1 | 0.1 | 0.2 | 0.1 | Yes |
| gpt_3_5_turbo_instruct | 15 Mar 2022 | 0.6 | 0.3 | 1.1 | 0.3 | Yes |
| gpt_4 | 14 Mar 2023 | 4.0 | 1.9 | 8.0 | 0.9 | Yes |
| gpt_4_1106_inspect | 6 Nov 2023 | 4.0 | 1.9 | 8.4 | 0.8 | Yes |
| claude_3_opus_inspect | 4 Mar 2024 | 4.0 | 1.7 | 8.8 | 0.6 | No |
| gpt_4_turbo_inspect | 9 Apr 2024 | 3.7 | 2.0 | 6.7 | 0.9 | No |
| gpt_4o_inspect | 13 May 2024 | 7.0 | 4.0 | 12.9 | 1.3 | Yes |
| claude_3_5_sonnet_20240620_inspect | 20 Jun 2024 | 11.4 | 5.5 | 22.4 | 1.7 | Yes |
| o1_preview | 12 Sep 2024 | 20.3 | 11.7 | 33.4 | 4.4 | Yes |
| claude_3_5_sonnet_20241022_inspect | 22 Oct 2024 | 20.5 | 10.1 | 40.8 | 2.6 | Yes |
| o1_inspect | 5 Dec 2024 | 38.8 | 21.2 | 65.0 | 7.1 | Yes |
| claude_3_7_sonnet_inspect | 24 Feb 2025 | 60.4 | 33.0 | 104.2 | 12.1 | Yes |
| o3_inspect | 16 Apr 2025 | 119.7 | 74.6 | 190.9 | 30.0 | Yes |
| claude_4_opus_inspect | 22 May 2025 | 100.4 | 60.0 | 163.5 | 20.4 | No |
| claude_4_1_opus_inspect | 5 Aug 2025 | 100.5 | 59.3 | 159.5 | 23.5 | No |
| gpt_5_2025_08_07_inspect | 7 Aug 2025 | 203.0 | 112.6 | 405.6 | 38.3 | Yes |
| gemini_3_pro | 18 Nov 2025 | 224.3 | 139.6 | 379.2 | 54.1 | Yes |
| gpt_5_1_codex_max_inspect | 19 Nov 2025 | 223.7 | 134.3 | 396.2 | 50.6 | No |
| claude_opus_4_5_inspect | 24 Nov 2025 | 293.0 | 161.7 | 623.7 | 49.4 | Yes |
| gpt_5_2 | 11 Dec 2025 | 352.2 | 198.1 | 815.2 | 66.0 | Yes |
| claude_opus_4_6_inspect | 5 Feb 2026 | 718.8 | 316.7 | 3,633.8 | 69.9 | Yes |
| gpt_5_3_codex | 5 Feb 2026 | 349.5 | 194.9 | 816.4 | 54.7 | No |
| gemini_3_1_pro | 19 Feb 2026 | 384.1 | 233.5 | 694.8 | 89.8 | No |
| gpt_5_4 | 5 Mar 2026 | 341.7 | 186.6 | 768.8 | 53.9 | No |
| claude_mythos_preview_early_inspect | 7 Apr 2026 | 1,044.8 | 508.9 | 3,304.3 | 185.9 | Yes |
Source: METR, Time Horizon v1.1 · as of 7 Apr 2026 · fetched 24 Oct 2026 · Download CSV (all columns)
Data pendingSelf-replication evaluation scores over time
There is no single maintained, machine-readable series yet. Published results are summarised below.
Published self-replication and scheming results (high level)
- Fudan University (Dec 2024): in agentic setups, systems driven by Llama 3.1-70B-Instruct and Qwen2.5-72B-Instruct created a live, separate copy of themselves in 50% and 90% of trials respectively (arXiv 2412.12140).
- UK AISI RepliBench (Apr 2025): across 20 task families (86 tasks), the five frontier models tested "do not currently pose a credible threat of self-replication, but succeed on many components". The best, Claude 3.7 Sonnet, scored >50% pass@10 on 15 of 20 families (arXiv 2504.18565).
- Apollo Research (Dec 2024): o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro and Llama 3.1 405B all showed in-context scheming when strongly goal-directed (arXiv 2412.04984).
- METR (Nov 2024) set out a threat model for "rogue replication" and the capabilities it would require (METR).
What to watch
- New METR time-horizon points for each frontier release.
- RepliBench and other replication evaluations in lab system cards.
- Lab responsible-scaling thresholds for autonomy.
Reading
- Time horizons— METR, 2026The live source for the chart.
- Measuring AI Ability to Complete Long Software Tasks— METR (Kwa et al.), 2025-03The method and the seven-month doubling.
- RepliBench— UK AI Security Institute, 2025-04Component-level replication evaluations.
- Frontier AI systems have surpassed the self-replicating red line— Fudan University, 2024-12The most-cited positive result.
- The Rogue Replication Threat Model— METR, 2024-11-12How replication risk is reasoned about.
Signals
Resilience and autonomy rising
- Time horizons keep doubling (chart).
- More replication components pass in evaluations (RepliBench).
- Multi-agent deployments with persistent memory become mainstream.
Contained
- Robust KYC, compute governance and weight security close off resource acquisition (RepliBench found models struggle with KYC).
- The horizon trend bends (time-horizon growth slows).
- Effective monitoring and oversight shown at scale.
What to watch
- Compute-governance and KYC rules for cloud providers.
Reading
- Frontier Models are Capable of In-context Scheming— Apollo Research, 2024-12Deception evaluations.
- International AI Safety Report— International AI Safety Report, 2026A consensus view on autonomy risks.
- An Overview of Catastrophic AI Risks— Hendrycks et al., 2023Includes "rogue AIs" and evolutionary pressures.
Prediction markets
Live prices from public prediction markets, snapshotted by scripts/update_markets.py. A price of 30% means traders pay about 30¢ for a contract that pays $1 if the event happens; treat it as a crowd estimate, not a fact. Polymarket and Kalshi are real-money markets (Kalshi is CFTC-regulated); Manifold uses play money and is often thin. Long-dated and low-volume markets are poorly calibrated. Each market's exact resolution rules on the platform decide what counts, so read them before relying on a headline question.
Self-replication and escape
83%
Ṁ3k play-money · 25 traders · closes 2030-01-01 · as of 2026-10-30 17:06 UTC
39%
Ṁ2k play-money · 30 traders · closes 2027-06-30 · as of 2026-10-30 17:06 UTC
24%
Ṁ8k play-money · 58 traders · closes 2027-03-31 · as of 2026-10-30 17:06 UTC
15%
$3k vol · $77 liquidity · closes 2026-11-01 · as of 2026-10-30 17:06 UTC
0%
$636 vol · closes 2026-10-16 · as of 2026-10-30 17:06 UTC
Agent capability and deployment
75%
$11k vol · $628 liquidity · closes 2026-12-01 · as of 2026-10-30 17:06 UTC
15%
$5k vol · $998 liquidity · closes 2026-12-01 · as of 2026-10-30 17:06 UTC
2% bid 2¢ / ask 3¢
66k contracts · 24k open interest · closes 2026-11-01 · as of 2026-10-30 17:06 UTC
8% bid 11¢ / ask 13¢
43k contracts · 13k open interest · closes 2026-12-01 · as of 2026-10-30 17:06 UTC
19% bid 16¢ / ask 19¢
32k contracts · 7k open interest · closes 2027-01-01 · as of 2026-10-30 17:06 UTC
Kalshi Will OpenAI release a personal AI agent before Oct 1, 2026? · Before Oct 1, 2026 resolved: No
0%
128k contracts · 34k open interest · closes 2026-10-01 · as of 2026-10-30 17:06 UTC
1% bid 0¢ / ask 6¢
4k contracts · 2k open interest · closes 2026-11-01 · as of 2026-10-30 17:06 UTC
16% bid 11¢ / ask 16¢
7k contracts · 2k open interest · closes 2026-12-01 · as of 2026-10-30 17:06 UTC
24% bid 23¢ / ask 28¢
5k contracts · 3k open interest · closes 2027-01-01 · as of 2026-10-30 17:06 UTC
Kalshi Will Anthropic release a personal AI agent before Oct 1, 2026? · Before Oct 1, 2026 resolved: No
0%
327 contracts · 327 open interest · closes 2026-10-01 · as of 2026-10-30 17:06 UTC
Oversight
13% bid 5¢ / ask 13¢
1k contracts · 866 open interest · closes 2027-01-01 · as of 2026-10-30 17:06 UTC
8% bid 4¢ / ask 8¢
15k contracts · 11k open interest · closes 2027-01-01 · as of 2026-10-30 17:06 UTC
23% bid 24¢ / ask 32¢
139 contracts · 66 open interest · closes 2027-07-01 · as of 2026-10-30 17:06 UTC
48% bid 43¢ / ask 48¢
101 contracts · 28 open interest · closes 2028-01-01 · as of 2026-10-30 17:06 UTC
Snapshot fetched 2026-10-30 17:06 UTC from public APIs (Polymarket Gamma/CLOB, Kalshi trade API, Manifold v0). Prices are what traders are paying, not my forecasts and not advice; thin or long-dated markets can be far off. Sparklines show daily prices over up to 120 days. Metaculus: API now requires an authenticated account token (HTTP 403 without one) and the website is behind a bot challenge, so Metaculus forecasts are not included. Add a ref type with a personal token later if wanted.
How to add a market
Add a ref to "markets" → "groups" → "refs" in src/themes/<slug>.json: {"platform":"polymarket","event":"<slug after /event/ in the URL>"}, {"platform":"kalshi","event":"<EVENT_TICKER>"} or {"platform":"manifold","id":"<market id or slug>"}. Optional: "markets" (list of specific sub-market slugs or tickers), "max", "sort" (date, prob or api). Then run python3 scripts/update_markets.py <slug> and rebuild.
Notes & journal
Dated observations from src/notes/agent-swarm-dna.md. Add a new entry with a ## YYYY-MM-DD Title heading.