Standing thesis
This document states the publication's standing argument and keeps the evidence attached: each claim links the field notes that raised it and the repositories that test it.
01 / Harness Engineering
For most startups, the model is no longer the only scarce layer. The scarce layer is the harness: context, constraints, evals, memory, cleanup loops, permissions, and product surfaces that make intelligence operable.
As foundation models become more capable and more accessible, early AI companies will compete on how well they package model capability into reliable workflows. This turns applied AI engineering into systems design, product taste, and operating discipline.
Field notes
Harness Engineering: the next battlefield for AI engineers
A high-signal X essay framing AI agents as model plus harness, with strong engagement around context engineering, architecture constraints, and entropy cleanup.
A real harness engineering project: persistent skill memory
Public post explaining why coding agents repeat mistakes and why per-skill memory is an operating layer, not just storage.
Lab proofs
fireworks-skill-memory
Proof: Turns repeated agent mistakes into recoverable operating context.
fireworks-sessions-saver
Proof: Preserves enough session structure for interruption, recovery, and handoff.
Founder use
02 / Long-Running Agents
Short demos hide the hard parts. Real agentic products need continuation, permission design, state recovery, subtask delegation, and progress that can be inspected after hours of work.
The winning agent products will feel less like chat sessions and more like operating systems for work. They need durable state, resumable intent, and fewer pointless interruptions.
Field notes
What is the limit of long-running agent tasks?
A field report after running Codex goals for many hours, turning usage pain into a product thesis around context compression, memory, and fit-for-purpose autonomy.
Codex goals unlock long tasks, delegation, and async work
A current X note positioning goals as a product primitive for long-running agent work, not just another coding feature.
The real pain is interruption, not capability
A widely saved post arguing that agent interruptions break the user's thinking state, which is a product-design problem for AI tools.
Lab proofs
fireworks-sessions-saver
Proof: Frames interruption and recovery as first-class product surfaces.
fireworks-ai-tools-memory
Proof: Explores cross-session tool memory for practical agent work.
Founder use
03 / Developer Communication
In AI infrastructure and devtools, explanation is part of the product. Diagrams, docs, launch examples, and reusable artifacts help a team sell trust before the product is fully obvious.
A founder who can explain a system repeatedly without redrawing it from scratch has a distribution advantage. Technical communication becomes onboarding, sales engineering, developer education, and investor clarity at once.
Field notes
fireworks-tech-graph launch post
The strongest public signal: high engagement around the pain of turning clear architecture intent into beautiful, reusable technical diagrams.
fireworks-tech-graph v3 update
A follow-up showing fast iteration from community feedback: Codex compatibility, output stability, and publication-quality examples.
Major upgrade from user feedback
Proof that the repo was not a one-off launch, but a feedback-driven product surface.
Lab proofs
Founder use
04 / Skill-Native Workbench
The most interesting AI workbench is not a single giant app. It is a set of small, inspectable capabilities that agents can call, update, combine, and hand off across contexts.
This points toward a new product category: AI-native workbenches where skills are distribution units, not hidden implementation details. The developer experience matters as much as the model.
Field notes
skills-updater and recommendation workflows
Public interest around keeping local AI skills updated and discoverable.
A practical example of turning media browsing, summarization, and screenshots into one repeatable agent workflow.
A workflow that removes manual asset search and download steps for creators and builders.
A recent example of packaging ambient coding context as a CLI-first AI-native workflow.
Lab proofs
fireworks-radio
Proof: Explores AI-native ambient workflows for coding sessions.
fireworks-tech-graph
Proof: Shows how one skill can become a public developer tool.
Founder use
05 / Retrieval and Skill-RAG
Retrieval systems fail when questions, evidence, memory, and tool actions are not aligned. The interesting work is diagnosing failure modes and turning them into skills, routing, and better product loops.
Many AI products are really search, routing, and evidence products with a model on top. Teams that understand this will build more trustworthy systems than teams that only tune generation.
Field notes
A note connecting adaptive retrieval failure to reusable skills and diagnostic workflows.
Dynamic software and agent infrastructure
A market read on deterministic execution, context as state, long sessions, tracing, self-checks, and approval mechanisms.
Lab proofs
fireworks-skill-memory
Proof: Turns lessons from failure into reusable memory lanes.
paper-pal
Proof: Frames research reading as an active synthesis and return workflow.
Founder use
06 / Distribution Systems
The same systems thinking that makes AI workflows reliable can also make public distribution compound: programmatic SEO, technical case studies, X field notes, GitHub metadata, and founder outreach all reinforce one another.
Early AI startups often under-package their knowledge. A durable site, searchable artifacts, and public proof loops can turn technical work into opportunity flow.
Field notes
Do not confuse impressions with traffic
A current note studying programmatic SEO and explaining why stable search growth beats short-term social feedback.
Lab proofs
bradzhang.dev
Proof: This site acts as a live distribution system for X, GitHub, SEO, and paid collaboration.
fireworks-tech-graph
Proof: A concrete example of turning OSS traction into a commercial founder-facing case.
Founder use