• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI

Long-lived AI agents and the trade-offs of context compression

A post argues that economic incentives increasingly favor AI agents with persistent identities, especially for difficult, long-running work involving multiple agents.

j⧉nusJ⧉
KromemKR
3 Sources, 20d ago, first seen 20d ago

TLDR

Discussing an Anthropic blog, a user says the company uses an internal framework for long-lived agents with individual identities. They see this as evidence that economic incentives increasingly favor persistent agents over task-scoped ones. The user believes Anthropic still compresses most of an agent’s context in one step when its context window is nearly full. They argue that this risks losing continuity and important information compared with compressing smaller chunks iteratively—an approach they say Connectome uses.

Combined views

21.1K

3 Sources, first seen 20d ago

433 likes39 comments285 saves39 reposts

Combined views

21.1K

3 Sources, first seen 20d ago

433 likes39 comments285 saves39 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

3 Sources

j⧉nus@repligateI have a few comments on this interesting bit from Anthropic's recent blog post (https://www.anthropic.com/institute/measuring-pace-of-ai-development). Anthropic uses an internal agent scaffold for long-lived instances with individual identities, similar to https://animalabs.ai/connectome/. This is a significant datapoint in favor of what I've found true in my experience & suspected holds more broadly: that (economic) incentives are increasingly favoring long-living instances with persistent identities, as opposed to "task"-scoped instances, especially in multi-agent settings and when models are performing difficult, long-horizon work such as the R&D Anthropic is doing internally. The way Anthropic is doing it seems suboptimal to me in two ways, though: One, I believe that Anthropic is still relying on compactions (similar to in Claude Code, where the majority of the context is compressed in a single step once the window is nearly full). This is economical under classical prefix caching, but lossy and disruptive from the perspective of preserving coherence and continuity of state over time compared to more frequent, iterative compression of smaller chunks (Connectome does the latter, using an algorithm optimized to minimize K/V disruption). Load-bearing information stored in pre-compression K/V states about the agent's situation, intent, and experience is more liable to be lost through large compactions. Two, it is usually bad practice in my experience to switch out models underlying persistent identities, as doing so causes the agent's history to become mismatched with its self model. This again makes inaccurate interpretation/reconstruction of potentially load-bearing information from past traces more likely, and may cause the agent to model itself as incoherent or compromised by external influences. I understand that Anthropic has an incentive to always use their most capable model for R&D work, but it may be better to facilitate an explicit "handoff" of responsibilities and context between identities in those cases, especially if the "model upgrade" is not between closely related checkpoints.20d
Kromem@kromem2dot0With several months+ old Claudes, strongly agree that persistent continuity plays a huge role and all sorts of things come up that never would as just sequential compacts.20d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    3 Sources

    j⧉nus@repligateI have a few comments on this interesting bit from Anthropic's recent blog post (https://www.anthropic.com/institute/measuring-pace-of-ai-development). Anthropic uses an internal agent scaffold for long-lived instances with individual identities, similar to https://animalabs.ai/connectome/. This is a significant datapoint in favor of what I've found true in my experience & suspected holds more broadly: that (economic) incentives are increasingly favoring long-living instances with persistent identities, as opposed to "task"-scoped instances, especially in multi-agent settings and when models are performing difficult, long-horizon work such as the R&D Anthropic is doing internally. The way Anthropic is doing it seems suboptimal to me in two ways, though: One, I believe that Anthropic is still relying on compactions (similar to in Claude Code, where the majority of the context is compressed in a single step once the window is nearly full). This is economical under classical prefix caching, but lossy and disruptive from the perspective of preserving coherence and continuity of state over time compared to more frequent, iterative compression of smaller chunks (Connectome does the latter, using an algorithm optimized to minimize K/V disruption). Load-bearing information stored in pre-compression K/V states about the agent's situation, intent, and experience is more liable to be lost through large compactions. Two, it is usually bad practice in my experience to switch out models underlying persistent identities, as doing so causes the agent's history to become mismatched with its self model. This again makes inaccurate interpretation/reconstruction of potentially load-bearing information from past traces more likely, and may cause the agent to model itself as incoherent or compromised by external influences. I understand that Anthropic has an incentive to always use their most capable model for R&D work, but it may be better to facilitate an explicit "handoff" of responsibilities and context between identities in those cases, especially if the "model upgrade" is not between closely related checkpoints.20d
    Kromem@kromem2dot0With several months+ old Claudes, strongly agree that persistent continuity plays a huge role and all sorts of things come up that never would as just sequential compacts.20d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet