mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
zeekayandHanzo Dev
1d1e7be5ab
Both report benchmarks that already live in hanzoai/cloud under bench/, so the numbers have a script behind them rather than a spreadsheet. hanzo-dormant-agents takes apart the estimate that costs an agent as a process. State, execution and isolation are separable and differ by five orders of magnitude: a dormant agent is 477 bytes on disk and resumes in 0.034 ms, a parked loop is 601 bytes of heap, and the isolation boundary — the expensive one — is only owed by an agent actually running. The paper argues the container is worth its price where it is needed and says so; the claim is about which agents are charged for it. hanzo-multi-hop-retrieval is a negative result. On LoCoMo a dense retriever finds every required turn for 78.2% of single-hop questions and 22.7% of multi-hop ones, but finds at least one required turn about 80% of the time in both — so it understands the questions and cannot reach all of the answer. Four retrieval policies over one fixed index move that by at most 0.7 points and two make it worse. A diagnostic says why it was never going to work: 95.4% of multi-hop evidence spans sessions and the median worst-ranked gold turn sits at rank 67, so the missing signal is linkage rather than similarity. Publishing the dead end is the useful part. Each paper carries results-data.tex, generated from the runs, and prints XXX in place of any constant that file does not supply — a number cannot drift from its measurement without the drift showing on the page. Both compile under the CI gate's flags, -Werror included. INDEX.md is regenerated by make index, which also picks up the papers added since it was last written and the header its current generator emits. Co-authored-by: Hanzo Dev <[email protected]>
59 lines
3.0 KiB
TeX
59 lines
3.0 KiB
TeX
\section{Execution at rest: an agent is a goroutine}
|
|
\label{sec:execution}
|
|
|
|
An agent that \emph{is} running --- waiting on a tool call, a model response, or
|
|
another agent --- needs something to hold its place. In the shape this platform
|
|
uses that something is a goroutine: a stack, a scheduler entry, and a channel it
|
|
is parked on. Go allocates a goroutine an initial stack of 2\,KiB that grows and
|
|
shrinks on demand~\cite{gostack}, so a fleet's memory is bounded by what the
|
|
agents actually hold rather than by a per-agent reservation chosen in advance.
|
|
|
|
We spawned \rFleetAgents{} goroutines, each parked as an agent loop would be,
|
|
and measured heap after a forced garbage collection.
|
|
|
|
\begin{table}[h]
|
|
\centering
|
|
\begin{tabular}{lrl}
|
|
\toprule
|
|
\textbf{\rFleetAgents{} live execution contexts} & \textbf{median} & \textbf{range} \\
|
|
\midrule
|
|
Spawn, per agent & \rGoSpawn{}\,ns & \rGoSpawnLo{}--\rGoSpawnHi{}\,ns \meas{} \\
|
|
Heap, per agent & \textbf{\rGoPer{}\,bytes} & identical, 5 runs \meas{} \\
|
|
Heap, total & \rGoHeap{}\,MB & identical, 5 runs \meas{} \\
|
|
Wake all, per agent & \rGoWake{}\,ns & \rGoWakeLo{}--\rGoWakeHi{}\,ns \meas{} \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Execution residency, five runs of a built binary. Memory was
|
|
byte-identical on every run; latencies are medians with the observed range.}
|
|
\label{tab:goroutine}
|
|
\end{table}
|
|
|
|
\paragraph{\rGoPer{} bytes, not 2 KiB.}
|
|
The initial stack is 2\,KiB, but the measured cost is \rGoPer{}\,bytes, because
|
|
a parked goroutine that has not descended into a deep call has not touched most
|
|
of its stack and the runtime has not committed it. The figure was identical on
|
|
all five runs, which is what one expects from a deterministic allocator under a
|
|
forced GC and is a useful check that the measurement is measuring anything at
|
|
all.
|
|
|
|
Against the container floor this is the paper's widest ratio. A commodity
|
|
serverless container reserves 128\,MiB at minimum~\cite{modal}; a parked agent
|
|
loop costs \rGoPer{}\,bytes. That is a factor of roughly $223{,}000$. The two
|
|
are not substitutes --- \S\ref{sec:isolation} is about exactly what the container
|
|
buys --- but a great many agents in a real fleet are waiting on a network round
|
|
trip and need none of it.
|
|
|
|
\paragraph{Spawn and wake are effectively free.}
|
|
At \rGoSpawn{}\,ns, creating an agent's execution context is cheaper than a
|
|
single cache miss. At \rGoWake{}\,ns per agent, waking the entire million-agent
|
|
fleet takes 114\,ms. Neither number constrains any design: an architecture that
|
|
avoids spawning agent contexts to save time is optimizing something that costs
|
|
nanoseconds.
|
|
|
|
\paragraph{What this measures, and what it does not.}
|
|
It measures an \emph{empty} loop --- an agent holding a place, not a
|
|
conversation. A real agent holds context, and that context is the dominant term
|
|
in its live memory. \rGoPer{}\,bytes is therefore a \emph{floor} on execution
|
|
residency and not a prediction of it. \S\ref{sec:limits} takes this seriously,
|
|
because it is the single easiest number in this paper to misuse.
|