The Log is the Agent
arxiv.org
arxiv.org
Event graphs aren’t just the agentic foundation for $MY_HARNESS — they’re the working cognitive substrate, native to what our favorite toolcall gremlins actually consume.
(Looking for lead investors for our angel syndicate btw! DM me if interested)
Reading your comment reminded me; I actually did something quite similar at $MY_BETTER_STARTUP! My approach is slightly different, however, employing what I like to call State-Horizon-Aware-Rercursive-Threaded-Graph-Position-Topology.
With a 400% increase in words, $MY_WAY_BETTER_HARNESS looks to be about four times as performant. SHARTGPT isn’t just a harness engineer's playground — it's a the cyber jungle gym that frees them from $MY_HARNESS.
If you want, I can even include a sentence or two that will really tell those potential investors why they should shower you with money instead of the other commenter! Just say the word!
(Looking for lead investors for our angel syndicate btw! DM me if interested)
Agentic development tends to encourage siloed individualistic development, so a lot of engineers reinvent similar patterns from first principles. It’s easier to write your own new thing than survey other approaches, so you’re more likely to perceive good ideas as original to your session.
This thread is helping me see how others addressed the same ideas.
I get and share criticism in the context of “AI hype startup culture”.
But out of it, this thread is really valuable for me mostly because of the links to codebases.
“Siloed individualistic development” openly shared is the part of HN I like most.
In a sense, sharing products in threads like this helps future people see what else has been done!
i had asked about this a while ago
All relevant events that affect the context window are stored in an event log. Forking agents and sessions is simply setting a pointer to the sequence number of another event log.
So if you want to check an implementation of this pattern see: https://github.com/smartcomputer-ai/lightspeed
In Lightspeed, we store all of them as events, thus can always reconstruct the exact state of the loop (e.g. state of open tool calls, compaction decisions in-flight, etc). This makes it possible to run the agent in a durable workflow engine easily.
But what I tried to get at is the question what additional events you store to construct more than just the llm session log but also more fine grained events around the entire agent state, which of course depends on what you want out of your agent.
The paper here in question is going even further and is event sourcing a larger state than just the session transcript, specifically additional graph structures that are getting built as part of the session.
in my agent, specifically, I focus on the event sourcing all the stuff that makes an agent work well as part of a deterministic workflow, which again is prerequisite to run agents in durable workflow engines like Temporal.
I write more about my approach here: https://github.com/smartcomputer-ai/lightspeed/blob/main/doc...
Another critique: the abstract mentions how their system allows for “branch[ing] a run at any event without re-executing the shared prefix,” but that’s only possible with very careful KV caching. Generally, rerunning inference from an earlier point still incurs O(n) input token cost and this paper is working at the wrong layer to see that. In this work, execution refers to tool calls but token generation is the expensive part.
It's more like the log is the only user/agent accepted consensus. It has to be the grounding base. Although extending it into an agentic system architecture becomes something not necessarily effective in practice.
That said, there are a _lot_ of "logs for agents" papers that I've read (and unfortunately gotten assigned to review) which are basically "we asked claude to hack on a graph DB and generate a paper".
[0] https://onewill.ai/blog/2026/stealing-50-years-of-database-i...
> rr records a group of Linux user-space processes and captures all inputs to those processes from the kernel, plus any nondeterministic CPU effects performed by those processes (of which there are very few).
I think the solution will resemble that. You don't control the LLM, sure. But you can control what it sees, and maybe that's good enough.
The paper’s pip library can be tried here
I’ve come to the same conclusion building my own agents. It simply feels ‘wrong’ that most frameworks will happily mutate your context. You have to explicitly go out of your way to store the original events. I’ve now started storing an event log for my own agents, this is used as the source of truth for deriving all subsequent context.
The great thing about this is that I have finer control over drift in long runs, as I can look back through the conversation/tool history and build context suitable for the current state of the agent. It also allows me to run compactions across the entire event history instead of ‘compactions on top of compactions’ which happens on long runs with checkpoints.
It definitely feels like this will be a bigger issue going forward as we have agents running longer and more complex workflows, I’ve started building a product aimed at addressing this issue in a framework agnostic way. [0]
There are pros and cons both ways.
NPS is more succinct and devoid of intermediate deviations, thus less likely to throw the agent off target.
ES otoh does contain deviations which might be salient in the future reasoning. If nothing else then to prune approaches that are known to be dead ends, but also fruitful approaches might have use especially when couple with the line of reasoning that lead to their application.
The problem I have with graph representations relative to LLMs is that you can never directly apply the concept. Everything that speaks to an LLM must ultimately be serialized. There's no getting around the token stream semantics.
I've found that one big flat markdown file tends to outperform everything. You could certainly project an event log into a graph and then serialize that, but it starts to feel like a Rube Goldberg machine at this point. It's a lot easier if you just work with the same terms that the models do.
Remember if that big document rarely changes and everything that comes before it is also constant, you'll pay something like 10% of the normal rate with providers like OAI for the tokens in those documents. The clever schemes to piecemeal out information feel good to the ego and might appeal to accounting at first glance, but I think the bitter lesson will ultimately win out here. We already have a million token context windows. Even if 90% of that is bullshit it's still a lot of tokens to work with.
"...and how it extends the BabyAGI lineage and prior graph-memory research. "
From BabyAGI from two years ago: "This is a framework built by Yohei who has never held a job as a developer. The purpose of this repo is to share ideas and spark discussion and for experienced devs to play with. Not meant for production use. Use with cautioun."
Chatbot is the command line
Agent is the bash script
___ is the GUI (macOS/Windows/GTA 6)
You need Xerox PARC all over again and we have one
but wouldn't feeding that log for each request/response iteration must get expensive really fast no?
also "We discuss--without claiming to demonstrate--" wtf? someone had a showerthought and slopped this out in 10mins to see what others thought?
The window on back-of-napkin-idea acquihires is closing fast. ;-)