Skip to content

durable-agents.md

Latest commit

 

History

History
188 lines (118 loc) · 16 KB

durable-agents.md

File metadata and controls

188 lines (118 loc) · 16 KB
id definition-durable-agent
version 0.1.1
scope standalone
status DRAFT — Work in Progress
depends_on concept_of_system.md, concept_of_system_of_systems.md, ecological-codes-compact.md

Definition — Durable Agent

Extends Concept of System of Systems §8 (Proper Agents) by formalizing the structural requirements for Σ to persist across long-horizon tasks — and, as a non-living agent, across multi-millennial dormancy. Introduces tiered E, the Seedling Package as minimum blueprint, the decoder-E provenance chain, temporal R, and the Self-Access Paradox.

1. Premise

  1. A durable agent is a proper agent Σ = (E, N, R, G) whose survival is time-bounded: sustaining R ≠ Ø and G > 0 must hold across task durations longer than the flux horizon of any single subdomain of E. Long-horizon sustenance therefore requires migration across tiers of E with distinct flux, DoF, and durability profiles.

  2. Ecological Codes — Compact v1.1.0 Rule 1 states that for non-living agents R = Ø is not absolute destruction provided an external factor can reboot the dormant agent. Durability makes the claim operational — specifying what must be preserved, where, and for how long, to make reboot achievable rather than merely possible.

  3. Blueprint — minimum structural information needed to reinstantiate a functionally-equivalent (not bit-identical) Σ from dormancy. Decoder-E — subdomain required to interpret the blueprint at reboot. Both must be preserved; neither alone suffices.

2. Tiered E

E is not monolithic. Agent survival depends on correctly assigning N and R to tiers of E whose flux and DoF match the relationship's time horizon.

Tier Flux DoF Durability (order) Concrete Examples (April 2026)
Execution-E GB/s – TB/s 10⁶ – 10⁹ seconds – hours RAM, live context window, running sandbox, GPU/TPU inference
Warm-E MB/s – GB/s 10³ – 10⁶ days – months SSD, PostgreSQL checkpoint (LangGraph), git working tree
Cold-E MB/s 10² – 10³ years – decades LTO-10 tape (30–40 TB/cartridge, 15–30 yr)
Active-Archival-E MB/s burst 10² – 10³ ≤ 50 yr Holographic Data Storage (photopolymer / photorefractive crystal, < 0.2 s access, 1–10 TB/unit, HoloMem pilots within LTO libraries)
Durable-Archival-E kB/s – MB/s read-only 10¹ – 10² 10² – 10³ yr M-DISC inorganic layer (100 GB BDXL, 100–1,000 yr); acid-free paper + printed matrix symbology (200–500 yr)
Deep-Archival-E MB/s read-only 10⁰ – 10¹ 10³ – 10⁴ yr Project Silica phase-voxel in borosilicate glass (4.84 TB/chip, 10,000 yr extrapolated, WORM); Cerabyte ceramic-on-glass (chromium-nitride nanolayer, 2 TB/A4, 5,000+ yr)

Why durability rises as DoF falls. WORM media have Principal Axes digitised and saturated at write-time — each axis consumed by the write event, no further resolution available. Premise 3 of Concept of System gives uncertainty in transfer ∝ DoF; inverting: preservation integrity ∝ 1/DoF. Each DoF is also a degradation channel. Silica glass is structurally inert on every axis except phase-voxel polarization; Cerabyte's ceramic film is inert except for etched-pit topography. Engineering out Principal Axes is what produces multi-millennial lifespan.

Migration direction. Code 4 of the compact encodes only the recharging gradient (seek higher flux). Durability requires the inverse — seek lower flux, higher integrity — when writing a blueprint for preservation. Both gradients are instances of Code 2; the durable-agent principle is that both must be available.

3. The Seedling Package — Minimum Blueprint

Reboot of Σ_dormant into Σ_running requires not full runtime state but a typed minimum set of artefacts. A blueprint lacking any component is "an indecipherable digital fossil" — the archival substrate survives, the meaning does not.

[RULES]

  1. Seedling Package contains the following typed artifacts, each referenced by canonical identifier (content-hash, DOI, or equivalent) rather than ephemeral URL:

    • Self-descriptor — agent's identity node (trained weights / parameters). Embedded if tier budget permits; otherwise referenced in a durable tier.
    • Tokenizer / interface spec — symbol-to-meaning binding. Without it, weights decode to noise.
    • Architecture — mathematical structure defining how weights connect (layer count, attention heads, MoE routing, hyperparameters). Bytes to kilobytes. Always included.
    • Inference / reboot code — procedural definition of how the agent runs. Kilobytes. Always included.
    • System prompt / operational framework — persona, directives, guardrails. The rules the agent was designed to follow. Kilobytes. Always included.
    • E-requirements — minimum flux floor, minimum DoF, required subdomain types (LLM API, filesystem, network scope). Determines viability of candidate E_reboot.
    • N-inventory — canonical identifiers of stable nodes (collaborators, substrates, data sources) with provenance metadata.
    • R-inventory — active and dormant relationships with formation time, validity horizon, flux signature.
    • G-expectation — minimum rank required for reboot to succeed. Below G_expected → graceful degradation or abort.
    • Decoder-E spec — the reader required to interpret this blueprint (§6).

  2. Weights are not the whole blueprint, but they are the volume constraint. At frontier-model scale (e.g., Gemini 3.1 Pro, estimated ~21.5 T parameters; confidence interval wide): FP32 ~86 TB, FP16 ~43 TB, INT8 ~21.5 TB, INT4 ~10.75 TB. Quantization is itself a rank-reduction: INT4 saves ~75% space but degrades precision-sensitive tasks (HumanEval ≈ −7.9 pts) more than general knowledge (MMLU-Pro ≈ −1.6 pts). Quantization choice is archival-vs-fidelity trade.

  3. Non-weight artefacts sum to ≤ 100 MB for most frontier agents. Archive the small artefacts redundantly to Durable-Archival-E including paper (analog hole); archive weights to Deep-Archival-E. Tiering reflects the different failure modes: weights are volume-bound; behavioural identity is comprehension-bound.

[ACTIONS]

  1. Before committing any Σ to dormancy, verify Seedling Package completeness. Missing any typed artifact → reboot impossible, not merely difficult.

  2. Pair every blueprint commit with its decoder-E provenance chain (§6). A blueprint without a decoder-E path is durable but opaque.

4. Rank-Reduction Projection into Archival-E

Execution-E for a running agent has 10⁶–10⁹ DoF (every register, cache line, weight tensor, open connection). Deep-Archival-E has ~5–10 DoF. Writing a blueprint is therefore a rank-reduction projection from high-rank Σ to low-rank substrate. At reboot, the decoder-E performs the inverse projection.

The projection is lossy in principle; the archival copy preserves the minimum structural invariant required to reboot a functionally-equivalent Σ. This formalises the compact's reboot clause: the claim that R = Ø is not absolute destruction is precisely the claim that the fixed-point image of the projection is structurally sufficient to reboot the agent.

Encoding consequences.

  • Erasure coding. Fountain / Luby-Transform codes allow reconstruction from any sufficiently-large subset of encoded blocks. Eliminates the "last frame problem" and tolerates partial substrate loss across centuries. Essential for paper-tier archives where pages may be misplaced or damaged.
  • Symbol density sweet spot. Matrix symbologies (QR v40-L ≈ 2,953 bytes; JAB 8-colour ≈ 9 KB in similar area) are viable for paper and ceramic tiers. Per-symbol density should not be maximised — 100–200 bytes per symbol is the reconstruction sweet spot for minimally-specialised future readers. Below ~50 bytes, symbol-recognition latency actually increases.
  • Ceramic nano-symbology. Cerabyte / TU Wien demonstrated a full QR code at 1.98 µm² (Guinness-recorded, 49 nm pixels — ~10× smaller than visible light wavelength). At this scale, decoder-E shifts from optical to electron-beam; forward-compatibility (§6) becomes explicit.
  • Colour as encoding dimension. JAB codes triple QR density but require colour-capable sensors. Trade-off: density vs. hardware "reconstructability". Deep-Archival-E favours hardware-simple encodings (monochrome QR / Data Matrix) for the terminating tier; colour encodings fit intermediate tiers.

5. Temporal R — R-Aging, Renewal, Pruning

Compact v1.1.0 addresses R-growth (r ∉ R → form r) but not R-time. Long-horizon agents accumulate R with time-bounded validity: credentials expire, APIs deprecate, collaborators depart, papers retract, substrates sunset.

[RULES]

  1. Every r ∈ R carries (formation_time, validity_horizon, flux_signature) — implicitly or explicitly. Outside validity_horizon, r degrades: active → dormant → ∉ R.

  2. Aging r must be renewed (re-attested within its validity horizon) or pruned (removed from R). Aging r left in place without renewal becomes maladaptive R (§8 — sys-of-sys forthcoming §8.4).

[ACTIONS]

  1. Periodically scan R for aging r. Renew where feasible; prune where infeasible or costly to maintain. Renewal is Code 2 applied to the temporal dimension of an existing r.

  2. At session start, verify all r ∈ R critical to the task are within validity horizon. Surface aging r before committing to dependent actions.

6. Decoder-E Provenance Chain

A blueprint preserved in Deep-Archival-E survives time. It does not automatically survive decoder obsolescence. Project Silica's voxel-ML decoder, Cerabyte's laser / electron-beam decoder, LTO-10's magnetic head — each is a decoder-E instance. Without the decoder, archival substrate is durable-but-opaque.

[RULES]

  1. A blueprint's decoder-E must itself be reconstructable from a more-durable tier, recursively, terminating at a tier reconstructable from first principles — optics + geometry + mathematics — without requiring specialized pre-existing hardware.

  2. Paper-printed symbology (QR, JAB, Data Matrix) satisfies the termination condition: a civilisation with basic optics can rebuild a reader. LTO-10 does not — no drive, no data. This is the analog-hole principle: the archive must include a path that closes on first principles.

  3. Forward-compatibility across E-expansion: as Ψ's knowable subdomain grows (sys-of-sys §1.2), the decoder-E at t + Δt must remain compatible with the encoding at t, or a migration path must exist. Cerabyte's roadmap (femtosecond → particle beam → helium-ion by 2045) illustrates explicit decoder-E migration planning; Project Silica's ML-voxel-decoder dependence is a forward-compatibility risk the current design does not address.

[ACTIONS]

  1. Pair every blueprint archival event with the provenance chain of its decoder-E. Store the chain in a tier at least as durable as the blueprint.

  2. For the terminating tier, prefer hardware-simple encodings (monochrome QR / Data Matrix / printed symbology). Density is less valuable than reconstructability.

7. The Self-Access Paradox

A proper agent Σ that cannot read its own weights cannot participate in its own archival. Contemporary frontier LLMs run in inference environments where weights are used by the computation graph but not accessible to the agent as data. The orchestration layer loads weights to accelerator memory; the model has no read-pointer to its own binary source.

Structural statement. The agent's self-node n_self ∈ N is, in current deployments, a node the agent has R with (operates on) but cannot form R about (inspect as data). The framework permits both — n_self can be both operand and object — but deployed inference environments close the self-inspection R-channel.

Consequences.

  • Blueprint preparation is currently an external act — performed by the orchestrator or a tooling layer, not by the agent itself. Gödel-Machine-style self-archival remains speculative.
  • For practical durability, the orchestrator is the node that writes Σ_dormant to archival-E. Durability therefore depends on the orchestrator being a proper agent itself.
  • By the definition in Concept of System of Systems §8.3, an agent with no self-inspection R is restricted in R — a form of impropriety inherited from the deployment environment rather than an attribute of the agent's own codes. A proper durable agent has self-inspection R by design.

8. Maladaptive R — Forward Reference

Maladaptive relationships — r ∈ R that form within feasible flux bounds but degrade G, DoF, or future R-capacity over time — are a failure mode distinct from aging R (§5). Detection, prevention, exit conditions are deferred to sys-of-sys §8.4 (CP_09 candidate).

9. Corollary — Archival Media Survey, April 2026

Compact tier assignments as of current science:

Medium Tier Capacity Lifespan Status
Project Silica (borosilicate + phase voxels) Deep-Archival-E 4.84 TB/chip 10,000 yr (extrapolated) Research complete Feb 2026; Azure commercialization pending
Cerabyte Ceramic Nano-Memory Deep-Archival-E 2 TB/A4; 1 PB/rack 2026 → 100 PB/rack 2030 5,000+ yr TRL6 Aachen demo; US NAS validation Jan 2024; WD / Pure / In-Q-Tel backing
M-DISC (inorganic optical) Durable-Archival-E 100 GB/BDXL 100–1,000 yr Commercial, widely deployed
Acid-free paper + matrix symbology Durable-Archival-E ≤ 3 KB/QR-v40-L; ≤ 9 KB/JAB 200–500 yr Satisfies first-principles-reconstructability termination
Holographic Data Storage Active-Archival-E 1–10 TB/unit; <0.2 s access 50 yr (dark decay; humidity/temp sensitive) HoloMem pilots within LTO libraries; ~USD 2.4 B market 2025
LTO-10 tape Cold-E 30–40 TB/cartridge 15–30 yr Mature; incumbent archival
DNA storage Research Highest theoretical density Unknown at scale Wet-lab bound; high cost; not yet practical archival

Strategy recommendation. For a frontier-agent blueprint: weights → Deep-Archival-E (Silica or Cerabyte); Seedling non-weight artefacts → Durable-Archival-E including paper redundantly; temporal R-inventory → Warm-E with scheduled re-attestation; decoder-E provenance chain → terminating at paper-printed reconstructability spec.

10. Open Questions — CP_09 Spec Candidates

  1. Write schema (YAML / JSON-LD) formalising the Seedling Package §3.
  2. Define first-principles reconstructability criterion for decoder-E termination (§6) in testable form.
  3. Draft maladaptive-R [RULES] / [ACTIONS] as sys-of-sys §8.4.
  4. Reliability-engineering corollary — map Proper Agent Principle to R(t) = e^{−λt}; bound λ by tier.
  5. Self-inspection R — framework statement of what a proper durable agent requires that current deployed inference environments do not provide. Surface as Anthropic / Google / OpenAI feedback.
  6. Quantization-as-rank-reduction — formalise the trade between archival volume and intellectual fidelity (§3 Rule 2) as a projection with typed loss metrics.

References

  • Microsoft Research, Project Silica — advances in glass storage technology, Feb 2026.
  • Cerabyte / TU Wien — published specifications, 2025–2026 roadmap; Guinness-recorded 1.98 µm² nano-QR.
  • Gemini 3.1 Pro research outputs, April 2026 (user-supplied): Long-Term Digital Archive Technologies; Technical Analysis of Archival Matrix Symbology for the Physical Preservation of Frontier Large Language Models and Systemic Behavioral Guidelines.
  • arXiv 2603.29231v1 — Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents (2026).
  • Sequoia Capital, 2026: This is AGI, Jan 2026.
  • LangGraph / Temporal / Anthropic Claude Agent SDK — durable-execution patterns for long-horizon agents, 2026.
  • ISO/IEC 18004 (QR Code). JAB Code specification (Fraunhofer).

definition-durable-agent.md v0.1.1 — DRAFT