Dispatching a Multi-Model Workforce from Anywhere
How the Agentic Driver extension set uses Herdr and Pi to route tasks by role, model and machine—and keep persistent sessions within reach from a laptop, phone or remote terminal.
Agent systems · models · evaluation
fieldnotes
Writing and evidence: agent systems, models, evaluation and computational biology.
Agent systems5 pieces Harnesses, gates, sandboxes, orchestration, and the products built on them.
Wrangling Qwen's Long Thinking Runs
How I manage Qwen's tendency to go off on a long reasoning run, why completion limits are not enough, and where quantisation creates a second serving problem.
And the Simpsons Already Did It
Why AI infrastructure keeps rediscovering established primitives, and how to distinguish useful standardisation from inflated novelty claims.
Where should an agent keep its memory?
Model state, session history and external retrieval belong to different parts of the agent system.
Why I started building a local model workforce
From finishing my PhD to keeping knowledge, model choice and authority in my own systems.
Models6 pieces Adapting models to a job, and serving them on hardware I own.
Is fast decoding the be all and end all?
Four serving recipes for GLM-5.3-Flash on two DGX Sparks. The recipe that answers fastest is not the one that generates fastest, and none of the four escapes the constraints that matter for real agent work.
Wrangling Qwen's Long Thinking Runs
How I manage Qwen's tendency to go off on a long reasoning run, why completion limits are not enough, and where quantisation creates a second serving problem.
Teaching a 4B model how I work
The project brief: a task-bound Implementer, what it learned, and what the coding tests showed.
Teaching a 4B model how I work: the technical report
Two starting checkpoints, executable training evidence, and fifteen paired coding-retention runs.
A better prompt is not a better model
Seven local models, 184 output-prediction problems, and a prompt change that helped some and hurt others.
Can local models work out what Python code returns?
Seven models over 100 CRUXEval-O problems, with prompt, parser and infrastructure failures kept visible.
Evaluation12 pieces Designing a study, running it, and reporting what it did and did not show.
60% of the time, it works every time
Is published research data ready to train models? I revisit four circadian ChIP-seq studies and what their raw data can support under consistent reanalysis.
Is fast decoding the be all and end all?
Four serving recipes for GLM-5.3-Flash on two DGX Sparks. The recipe that answers fastest is not the one that generates fastest, and none of the four escapes the constraints that matter for real agent work.
Circadian ChIP-seq reproducibility audit
A method reconstruction, sensitivity analysis and local ENCODE-equivalent comparison for public mouse liver circadian factor ChIP-seq. No tested condition reproduced both the deposited peak counts and the peak sets.
Which local model is right for the job?
Choosing a model for a specific job, and deciding what to change when it fails.
Is cooling part of the inference budget?
A 6.6 W fan, a 35 W desk fan, and the cost of keeping a model running.
Teaching a 4B model how I work: the technical report
Two starting checkpoints, executable training evidence, and fifteen paired coding-retention runs.
Can a reasoning manual give your small models a capability uplift?
Run 5: a small-model signal on a storage-budget trap, with a same-length sham control.
A better prompt is not a better model
Seven local models, 184 output-prediction problems, and a prompt change that helped some and hurt others.
Does the reasoning manual help Qwen catch mistakes?
Run 4: Qwen created a little headroom during calibration, but all three arms passed the follow-up test.
Can local models work out what Python code returns?
Seven models over 100 CRUXEval-O problems, with prompt, parser and infrastructure failures kept visible.
Does a reasoning manual improve how Anthropic's Sonnet tackles error detection?
Sonnet 5, three prompt conditions, three difficulty levels and 27 responses: testing error detection separately from presentation.
The reasoning manual changed the explanation, not the outcome
Run 2: holding the model constant separated presentation changes from trap detection.
Computational biology2 pieces Circadian genomics, phenome classification, and disease modelling.
60% of the time, it works every time
Is published research data ready to train models? I revisit four circadian ChIP-seq studies and what their raw data can support under consistent reanalysis.
Circadian ChIP-seq reproducibility audit
A method reconstruction, sensitivity analysis and local ENCODE-equivalent comparison for public mouse liver circadian factor ChIP-seq. No tested condition reproduced both the deposited peak counts and the peak sets.
60% of the time, it works every time
Is published research data ready to train models? I revisit four circadian ChIP-seq studies and what their raw data can support under consistent reanalysis.
Is fast decoding the be all and end all?
Four serving recipes for GLM-5.3-Flash on two DGX Sparks. The recipe that answers fastest is not the one that generates fastest, and none of the four escapes the constraints that matter for real agent work.
Dispatching a Multi-Model Workforce from Anywhere
How the Agentic Driver extension set uses Herdr and Pi to route tasks by role, model and machine—and keep persistent sessions within reach from a laptop, phone or remote terminal.
Wrangling Qwen's Long Thinking Runs
How I manage Qwen's tendency to go off on a long reasoning run, why completion limits are not enough, and where quantisation creates a second serving problem.
And the Simpsons Already Did It
Why AI infrastructure keeps rediscovering established primitives, and how to distinguish useful standardisation from inflated novelty claims.
Circadian ChIP-seq reproducibility audit
A method reconstruction, sensitivity analysis and local ENCODE-equivalent comparison for public mouse liver circadian factor ChIP-seq. No tested condition reproduced both the deposited peak counts and the peak sets.
Which local model is right for the job?
Choosing a model for a specific job, and deciding what to change when it fails.
Where should an agent keep its memory?
Model state, session history and external retrieval belong to different parts of the agent system.
Is cooling part of the inference budget?
A 6.6 W fan, a 35 W desk fan, and the cost of keeping a model running.
Why I started building a local model workforce
From finishing my PhD to keeping knowledge, model choice and authority in my own systems.
Teaching a 4B model how I work
The project brief: a task-bound Implementer, what it learned, and what the coding tests showed.
Teaching a 4B model how I work: the technical report
Two starting checkpoints, executable training evidence, and fifteen paired coding-retention runs.
Can a reasoning manual give your small models a capability uplift?
Run 5: a small-model signal on a storage-budget trap, with a same-length sham control.
A better prompt is not a better model
Seven local models, 184 output-prediction problems, and a prompt change that helped some and hurt others.
Does the reasoning manual help Qwen catch mistakes?
Run 4: Qwen created a little headroom during calibration, but all three arms passed the follow-up test.
Can local models work out what Python code returns?
Seven models over 100 CRUXEval-O problems, with prompt, parser and infrastructure failures kept visible.
Does a reasoning manual improve how Anthropic's Sonnet tackles error detection?
Sonnet 5, three prompt conditions, three difficulty levels and 27 responses: testing error detection separately from presentation.
The reasoning manual changed the explanation, not the outcome
Run 2: holding the model constant separated presentation changes from trap detection.