Draft — this page specifies behaviour that is not implemented. It may change without notice. Last reviewed 2026-09-12.

Performance budgets

Purpose

This page states what each part of the system is allowed to cost, and what gives way first when something exceeds it.

Most numbers here are targets, not measurements

The model latency figures in particular cannot be written until the model has been run on the reference hardware with a game in memory, which is open question 1 in model contract.

Every unvalidated figure is marked. Do not treat one as a specification until it has been measured and this page updated.

The governing constraint

The agent and the game share one machine and one graphics device, and the game was there first.

This is the framing for everything below. A perception pipeline that is a little slow costs the agent some responsiveness. A perception pipeline that starves the game of graphics memory costs the user their game, and an agent that makes a game stutter is uninstalled regardless of how well it plays.

The reflex tick

Rate: 20–30 Hz, so a budget of 33 ms at the lower bound.

Phase	Budget	Status
Frame arrival and pre-pass, on the graphics device	4 ms	Target
Signature read-back and change classification	0.5 ms	Target, NFR-PERC-001
Sensor evaluation	1 ms	Target
Executor step: sustained actions, reconciliation	0.5 ms	Target, NFR-ACT-001
Safety interlocks	0.1 ms	Target
Telemetry publish	0.2 ms	Target, NFR-OBS-001
Headroom	~26 ms

The headroom is deliberate and large. The reflex loop is the one path with a hard deadline, and it is the path that keeps a key held; anything that erodes it shows up in-game as stutter, which is exactly the symptom users blame on the agent.

Extraction work — text recognition, element detection, template matching — runs on a worker pool and lands asynchronously. It is never in this budget.

Model calls

Call	Target	Status
Tactical, small image cost, no extended reasoning	150–400 ms	Unvalidated
Deliberative, large image cost, extended reasoning	3–15 s	Unvalidated
Two-stage refinement, both passes	Under the deliberative budget	Unvalidated

NFR-MODEL-001 is written against the tactical figure, and the tactical cadence's rate follows from it. Until it is measured, the rate in the reasoning loop is an intention.

Neither call is on the reflex path. NFR-LOOP-001 requires the other cadences to keep running while either is in flight, which is what makes a fifteen-second deliberation invisible.

Graphics memory

The budget must hold all of this at once:

Consumer	Notes
The game	4–10 GiB depending on title and settings. Assume the high end.
Model weights	The dominant agent cost
Vision encoder	Loaded alongside the weights
Attention cache	Scales with context length; the main dial
Perception models	Text recognition and element detection, well under a gigabyte
Capture surfaces and pyramids	Small

On the reference hardware — 24 GiB — a game at the high end leaves under half for everything else, which is why context length rather than model size is the first thing reduced.

Degradation ladder

FR-MODEL-004. Available memory is polled once a second, and when the game's headroom shrinks:

1. Reduce context length. Cheapest, and the attention cache is the largest variable cost.
2. Reduce tactical image cost. Directly trades grounding accuracy for memory and latency.
3. Move the deliberative cadence to the processor. Viable because only a fraction of the chosen model's parameters are active per token — an estimated single-digit to low-double-digit tokens per second, which is acceptable for a call that fires twice a minute. Unvalidated.
4. Pause and tell the user.

The ladder is ordered by what the user notices last. A shorter context is invisible until a session runs long; a stuttering game is noticed within seconds.

Safety latency

NFR-SAFE-001.

Path	Budget	Status
Panic key to last release event	100 ms	Asserted, not measured
Foreground loss to release	One reflex tick	Target
User input detected to pause	One reflex tick	Target
Core exit to guardian release	250 ms	Target

The panic figure is the one that matters most and is the least substantiated. It needs measuring against hotkey delivery latency under load, and it is open question 1 in safety and limits.

Interface responsiveness

NFR-GUI-001.

Surface	Budget
Overlay annotation update	One reflex tick
Shell interaction response	100 ms
Shell live view	At least 15 frames per second

The shell performs no inference, capture or perception, so its responsiveness is a property of the process boundary rather than of tuning.

Disk

Consumer	Budget
Session recording, default	2 GiB per session, then oldest keyframes dropped
Retention, default	7 days
Logs	100 MiB, rotated

An unattended overnight session at full frame rate would fill a disk. The quota is a default, it is user-adjustable, and purging is a single action.

What is measured, and where

Budget	Method	Where
Reflex phases	Instrumented timings, 95th percentile over a session	Nightly, reference hardware
Model calls	Timed at the host boundary	Nightly, with a game running
Graphics memory	Polled, recorded per session	Every session
Panic latency	Instrumented from key press to last event	Nightly and manually
Game frame rate impact	Measured with and without the agent	Manual, per game

The last row is the one that decides whether the project is usable, and it is the only one that cannot be automated.

When a budget is exceeded

Budget	Response
Reflex tick	Log, drop optional work, then reduce the reflex rate
Tactical call	Reduce image cost, then reduce the tactical rate
Deliberative call	Reduce reasoning depth, then reduce context
Graphics memory	The ladder above
Panic latency	A defect. Not degraded; broken.
Disk	Drop oldest keyframes, then stop recording, then stop the session

Everything degrades except safety. A safety budget exceeded is a defect to fix, not a threshold to relax.

Open questions

1. Every model figure is unvalidated. This blocks NFR-MODEL-001 and the cadence rates.
2. The 100 ms panic budget is asserted. It may be unachievable through the ordinary hotkey path under load, in which case the mechanism needs changing rather than the number.
3. Whether the reflex rate should adapt to the game's own frame rate. A game at 30 fps does not need a 30 Hz agent.
4. Whether the graphics memory poll at one second is frequent enough. A game loading a level can allocate gigabytes faster than that.
5. What the acceptable frame-rate impact actually is. "Within a stated fraction" is in the vision and the fraction is not stated, which makes the success criterion unmeasurable as written.

Related decisions

ADR-0006 and ADR-0007 govern the model figures, and ADR-0024 the safety path. None are written, and the first two are blocked on measurements taken with a game running.