Conor Bronsdon
@conorbronsdon.com
Talking/writing about startups, AI, politics, & software development
🎙 ex-Host of Dev Interrupted, current host of Chain of Thought ⛓️, failed founder, & avid sports fan.
🎙 ex-Host of Dev Interrupted, current host of Chain of Thought ⛓️, failed founder, & avid sports fan.
Pinned
Conor Bronsdon
@conorbronsdon.com
· Aug 24
Thomson 1: The New $40M Legal AI Model | Thomson Reuters Joel Hron
Behind Thomson, the new legal AI model from Thomson Reuters, is a $...
www.youtube.com
Thomson Reuters launched their new Thomson model for legal AI today - and their CTO Joel Hron joined me to break down their training approach, providing legal context to LLMs, citation approach, and more 👇
youtu.be/CQpAVPSYPxI
youtu.be/CQpAVPSYPxI
Cisco's Jeetu Patel: AI agents are like teenagers: "They are profoundly intelligent, extremely resourceful, have no fear of consequence."
Full episode:
chainofthought.show/podcast/67-...
Full episode:
chainofthought.show/podcast/67-...
September 29, 2026 at 4:30 PM
Cisco's Jeetu Patel: AI agents are like teenagers: "They are profoundly intelligent, extremely resourceful, have no fear of consequence."
Full episode:
chainofthought.show/podcast/67-...
Full episode:
chainofthought.show/podcast/67-...
Half a million AI workflow runs in a month (GrowthX's count, 2025). Marcel Santilli's point is the expert's time: 100 articles, 89 of them not relevant.
chainofthought.show/podcast/35-...
chainofthought.show/podcast/35-...
September 29, 2026 at 3:00 PM
Half a million AI workflow runs in a month (GrowthX's count, 2025). Marcel Santilli's point is the expert's time: 100 articles, 89 of them not relevant.
chainofthought.show/podcast/35-...
chainofthought.show/podcast/35-...
No dedicated PMs at Genspark. A bugs agent triages user feedback, assigns it to a developer, and the developer's agents fix, test and report back.
Their co-founder sat down with me: chainofthought.show/podcast/73-...
Their co-founder sat down with me: chainofthought.show/podcast/73-...
September 29, 2026 at 2:00 PM
No dedicated PMs at Genspark. A bugs agent triages user feedback, assigns it to a developer, and the developer's agents fix, test and report back.
Their co-founder sat down with me: chainofthought.show/podcast/73-...
Their co-founder sat down with me: chainofthought.show/podcast/73-...
Getting more from the GPUs you already have: AMD's Anush Elangovan joins me to explain how AI agents profile workloads and test optimizations with Hyperloom and ROCm 10, including while you sleep.
Full Chain of Thought episode: youtu.be/xMC7d_h5rSQ
Full Chain of Thought episode: youtu.be/xMC7d_h5rSQ
Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan
AI agents can profile GPU workloads, find slow kernels, and keep te...
www.youtube.com
September 29, 2026 at 1:30 PM
Getting more from the GPUs you already have: AMD's Anush Elangovan joins me to explain how AI agents profile workloads and test optimizations with Hyperloom and ROCm 10, including while you sleep.
Full Chain of Thought episode: youtu.be/xMC7d_h5rSQ
Full Chain of Thought episode: youtu.be/xMC7d_h5rSQ
Context OS v0.14.0 is out ✅
- OpenCode joins Claude Code, Codex, and OpenClaw as a 1st class host
- Hermes, Cursor, and Devin remain experimental for now (close!)
- Handoff previews now name their sources
- Local receipt history upgrades
github.com/conorbronsd...
Try it: github.com/conorbronsd...
- OpenCode joins Claude Code, Codex, and OpenClaw as a 1st class host
- Hermes, Cursor, and Devin remain experimental for now (close!)
- Handoff previews now name their sources
- Local receipt history upgrades
github.com/conorbronsd...
Try it: github.com/conorbronsd...
GitHub - conorbronsdon/agent-context-os: Portable, evolving context and workflow layer that works across Claude Code, Codex, Hermes Agent, OpenClaw, Cursor, Devin and other compatible coding agents. All backed by Git.
Portable, evolving context and workflow layer that works across Claude Code, Codex, Hermes Agent, OpenClaw, Cursor, Devin and other compatible coding agents. All backed by Git. - conorbronsdon/age...
github.com
September 29, 2026 at 5:36 AM
Context OS v0.14.0 is out ✅
- OpenCode joins Claude Code, Codex, and OpenClaw as a 1st class host
- Hermes, Cursor, and Devin remain experimental for now (close!)
- Handoff previews now name their sources
- Local receipt history upgrades
github.com/conorbronsd...
Try it: github.com/conorbronsd...
- OpenCode joins Claude Code, Codex, and OpenClaw as a 1st class host
- Hermes, Cursor, and Devin remain experimental for now (close!)
- Handoff previews now name their sources
- Local receipt history upgrades
github.com/conorbronsd...
Try it: github.com/conorbronsd...
LinkedIn not enabling people tags for non-native schedule posts is the bane of my existence
September 29, 2026 at 1:19 AM
LinkedIn not enabling people tags for non-native schedule posts is the bane of my existence
In critical infrastructure, your AI has to work across mainframes, older systems, and multiple clouds.
Kris Lovejoy explains why getting that data into usable shape is a deployment problem to solve early.
Watch the clip: youtube.com/shorts/xbvu...
Kris Lovejoy explains why getting that data into usable shape is a deployment problem to solve early.
Watch the clip: youtube.com/shorts/xbvu...
September 29, 2026 at 12:00 AM
In critical infrastructure, your AI has to work across mainframes, older systems, and multiple clouds.
Kris Lovejoy explains why getting that data into usable shape is a deployment problem to solve early.
Watch the clip: youtube.com/shorts/xbvu...
Kris Lovejoy explains why getting that data into usable shape is a deployment problem to solve early.
Watch the clip: youtube.com/shorts/xbvu...
avoid-ai-writing: an open-source skill that audits a draft for AI writing patterns and rewrites them. 74 pattern categories, works in Claude Code, Codex, Cursor and more.
Try it:
github.com/conorbronsd...
Try it:
github.com/conorbronsd...
September 28, 2026 at 8:55 PM
avoid-ai-writing: an open-source skill that audits a draft for AI writing patterns and rewrites them. 74 pattern categories, works in Claude Code, Codex, Cursor and more.
Try it:
github.com/conorbronsd...
Try it:
github.com/conorbronsd...
“Don’t delete the database” leaves a separate question: does the agent have permission to delete it?
Trace the action. What denies it? Can the agent change that rule or use another route?
My takeaway from Redpanda’s Tyler Akidau: enforce critical rules outside the agent.
Trace the action. What denies it? Can the agent change that rule or use another route?
My takeaway from Redpanda’s Tyler Akidau: enforce critical rules outside the agent.
September 28, 2026 at 7:45 PM
“Don’t delete the database” leaves a separate question: does the agent have permission to delete it?
Trace the action. What denies it? Can the agent change that rule or use another route?
My takeaway from Redpanda’s Tyler Akidau: enforce critical rules outside the agent.
Trace the action. What denies it? Can the agent change that rule or use another route?
My takeaway from Redpanda’s Tyler Akidau: enforce critical rules outside the agent.
Two models can agree because they share the same blind spot.
On Chain of Thought, Dan Klein explains correlated errors and the judgment it takes when AI turns you into the editor of its work.
chainofthought.show/podcast/54-...
On Chain of Thought, Dan Klein explains correlated errors and the judgment it takes when AI turns you into the editor of its work.
chainofthought.show/podcast/54-...
September 28, 2026 at 7:00 PM
Two models can agree because they share the same blind spot.
On Chain of Thought, Dan Klein explains correlated errors and the judgment it takes when AI turns you into the editor of its work.
chainofthought.show/podcast/54-...
On Chain of Thought, Dan Klein explains correlated errors and the judgment it takes when AI turns you into the editor of its work.
chainofthought.show/podcast/54-...
substack-mcp: an MCP server and CLI for Substack. Drafts, archive search, analytics, Notes, consented subscribers. Long-form posts are draft-only by design.
Try it:
github.com/conorbronsd...
Try it:
github.com/conorbronsd...
September 28, 2026 at 4:30 PM
substack-mcp: an MCP server and CLI for Substack. Drafts, archive search, analytics, Notes, consented subscribers. Long-form posts are draft-only by design.
Try it:
github.com/conorbronsd...
Try it:
github.com/conorbronsd...
AI Angels: a CC0 directory of 51 active angel investors in AI. Every entry has a public source and a last-verified date.
github.com/conorbronsd...
github.com/conorbronsd...
September 28, 2026 at 3:00 PM
AI Angels: a CC0 directory of 51 active angel investors in AI. Every entry has a public source and a last-verified date.
github.com/conorbronsd...
github.com/conorbronsd...
What is multimodal AI?
One model that reasons across text, images, audio and video. A hard part is alignment: connecting what it sees with what it reads.
Full breakdown:
chainofthought.show/ai-decoded/...
One model that reasons across text, images, audio and video. A hard part is alignment: connecting what it sees with what it reads.
Full breakdown:
chainofthought.show/ai-decoded/...
September 28, 2026 at 1:30 PM
What is multimodal AI?
One model that reasons across text, images, audio and video. A hard part is alignment: connecting what it sees with what it reads.
Full breakdown:
chainofthought.show/ai-decoded/...
One model that reasons across text, images, audio and video. A hard part is alignment: connecting what it sees with what it reads.
Full breakdown:
chainofthought.show/ai-decoded/...
Andy Masley: 1% of your daily emissions takes "about a thousand regular chatbot prompts." Now on TIME100 AI 2026.
chainofthought.show/podcast/45-...
chainofthought.show/podcast/45-...
September 28, 2026 at 2:14 AM
Andy Masley: 1% of your daily emissions takes "about a thousand regular chatbot prompts." Now on TIME100 AI 2026.
chainofthought.show/podcast/45-...
chainofthought.show/podcast/45-...
When should you use a small model instead of a frontier model?
Prototype big, serve small. Use the frontier to find out what "good" looks like, then move a stable, high-volume task to a smaller model tuned for it.
Full breakdown:
chainofthought.show/ai-decoded/...
Prototype big, serve small. Use the frontier to find out what "good" looks like, then move a stable, high-volume task to a smaller model tuned for it.
Full breakdown:
chainofthought.show/ai-decoded/...
September 28, 2026 at 12:00 AM
When should you use a small model instead of a frontier model?
Prototype big, serve small. Use the frontier to find out what "good" looks like, then move a stable, high-volume task to a smaller model tuned for it.
Full breakdown:
chainofthought.show/ai-decoded/...
Prototype big, serve small. Use the frontier to find out what "good" looks like, then move a stable, high-volume task to a smaller model tuned for it.
Full breakdown:
chainofthought.show/ai-decoded/...
What are AI agent guardrails?
Limits on what an agent can take in, do and output. A rule in the prompt is a request, not a control. Enforce the must-nevers in code, with permissions and approval gates.
Full breakdown:
chainofthought.show/ai-decoded/...
Limits on what an agent can take in, do and output. A rule in the prompt is a request, not a control. Enforce the must-nevers in code, with permissions and approval gates.
Full breakdown:
chainofthought.show/ai-decoded/...
September 27, 2026 at 10:00 PM
What are AI agent guardrails?
Limits on what an agent can take in, do and output. A rule in the prompt is a request, not a control. Enforce the must-nevers in code, with permissions and approval gates.
Full breakdown:
chainofthought.show/ai-decoded/...
Limits on what an agent can take in, do and output. A rule in the prompt is a request, not a control. Enforce the must-nevers in code, with permissions and approval gates.
Full breakdown:
chainofthought.show/ai-decoded/...
The systems analyst is coming back. npm co-founder Laurie Voss calls it the product engineer: figure out what users really need, then work with a team of agents to build it, without writing the code yourself.
chainofthought.show/podcast/74-...
chainofthought.show/podcast/74-...
September 27, 2026 at 5:00 PM
The systems analyst is coming back. npm co-founder Laurie Voss calls it the product engineer: figure out what users really need, then work with a team of agents to build it, without writing the code yourself.
chainofthought.show/podcast/74-...
chainofthought.show/podcast/74-...
podcast-benchmark: compare any podcast with a peer set on public signals (cadence, catalog, duration, transcripts, feed hygiene). Outputs JSON plus a ranked Markdown report.
github.com/conorbronsd...
github.com/conorbronsd...
September 27, 2026 at 4:00 PM
podcast-benchmark: compare any podcast with a peer set on public signals (cadence, catalog, duration, transcripts, feed hygiene). Outputs JSON plus a ranked Markdown report.
github.com/conorbronsd...
github.com/conorbronsd...
Why do multi-agent systems fail?
Usually in the gaps between agents. A handoff drops context, or one agent's wrong output becomes the next one's trusted input.
Treat the system as the unit: trace every step, validate every handoff.
Full breakdown:
chainofthought.show/ai-decoded/...
Usually in the gaps between agents. A handoff drops context, or one agent's wrong output becomes the next one's trusted input.
Treat the system as the unit: trace every step, validate every handoff.
Full breakdown:
chainofthought.show/ai-decoded/...
Why do multi-agent systems fail, and how do you make them reliable?
Multi-agent systems break in the handoffs between agents, where small faults cascade. How to trace, validate, and threat-model the whole system, from Chain of Thought's multi-agent and AI-security episodes.
chainofthought.show
September 27, 2026 at 3:00 PM
Why do multi-agent systems fail?
Usually in the gaps between agents. A handoff drops context, or one agent's wrong output becomes the next one's trusted input.
Treat the system as the unit: trace every step, validate every handoff.
Full breakdown:
chainofthought.show/ai-decoded/...
Usually in the gaps between agents. A handoff drops context, or one agent's wrong output becomes the next one's trusted input.
Treat the system as the unit: trace every step, validate every handoff.
Full breakdown:
chainofthought.show/ai-decoded/...
How do you test an AI system when the output isn't deterministic?
Test properties, not exact strings. Build a dataset of real inputs and what a good answer must do, score every output against it, and run it on every change like CI.
Full breakdown:
chainofthought.show/ai-decoded/...
Test properties, not exact strings. Build a dataset of real inputs and what a good answer must do, score every output against it, and run it on every change like CI.
Full breakdown:
chainofthought.show/ai-decoded/...
How do you test an AI system when the output isn't deterministic?
AI output isn't deterministic, so assert-equals tests don't work. How to test on properties and datasets instead, from Chain of Thought's episodes with Hamel Husain and Charity Majors.
chainofthought.show
September 27, 2026 at 1:09 AM
How do you test an AI system when the output isn't deterministic?
Test properties, not exact strings. Build a dataset of real inputs and what a good answer must do, score every output against it, and run it on every change like CI.
Full breakdown:
chainofthought.show/ai-decoded/...
Test properties, not exact strings. Build a dataset of real inputs and what a good answer must do, score every output against it, and run it on every change like CI.
Full breakdown:
chainofthought.show/ai-decoded/...
How do you cut the cost of running an AI agent?
Trace it first so you can see cost per step. Then right-size the model for each step, trim and cache context, and fix the failures that cause expensive retries.
Full breakdown:
chainofthought.show/ai-decoded/...
Trace it first so you can see cost per step. Then right-size the model for each step, trim and cache context, and fix the failures that cause expensive retries.
Full breakdown:
chainofthought.show/ai-decoded/...
How do you cut the cost of running an AI agent?
AI agent cost hides in redundant calls, oversized models, and bloated context. How to find and cut the real drivers, from Chain of Thought's episodes with Intercom and Superhuman.
chainofthought.show
September 26, 2026 at 11:27 PM
How do you cut the cost of running an AI agent?
Trace it first so you can see cost per step. Then right-size the model for each step, trim and cache context, and fix the failures that cause expensive retries.
Full breakdown:
chainofthought.show/ai-decoded/...
Trace it first so you can see cost per step. Then right-size the model for each step, trim and cache context, and fix the failures that cause expensive retries.
Full breakdown:
chainofthought.show/ai-decoded/...
Before an agent optimizes your workload: keep a baseline, define acceptable outputs, measure end-to-end, set a budget and keep a rollback path.
Faster code still needs a trustworthy test.
Faster code still needs a trustworthy test.
September 26, 2026 at 7:00 PM
Before an agent optimizes your workload: keep a baseline, define acceptable outputs, measure end-to-end, set a budget and keep a rollback path.
Faster code still needs a trustworthy test.
Faster code still needs a trustworthy test.
Test agent memory with a correction: change a deployment destination in the source, then start a fresh task.
Does the agent find the update or reuse the stale value?
Does the agent find the update or reuse the stale value?
September 25, 2026 at 10:45 PM
Test agent memory with a correction: change a deployment destination in the source, then start a fresh task.
Does the agent find the update or reuse the stale value?
Does the agent find the update or reuse the stale value?
An agent skill can go stale while your product keeps shipping. AMD's Anush Elangovan pointed to ROCm's open skills, docs, code and debugging context - and how they accelerate readiness through openness: chainofthought.show/podcast/72-...
Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan on ROCm 10 — Chain of Thought
AMD’s Anush Elangovan explains how AI agents profile GPU workloads, tune kernels, and keep searching for better performance with Hyperloom and ROCm 10. He…
chainofthought.show
September 25, 2026 at 7:00 PM
An agent skill can go stale while your product keeps shipping. AMD's Anush Elangovan pointed to ROCm's open skills, docs, code and debugging context - and how they accelerate readiness through openness: chainofthought.show/podcast/72-...
"Meeting notes are just start points of work. It's not the end point." Wen Sang on why Genspark's recorder turns a meeting into the proposal, the pricing model and the follow-ups instead of stopping at the summary.
chainofthought.show/podcast/73-...
chainofthought.show/podcast/73-...
September 25, 2026 at 3:30 PM
"Meeting notes are just start points of work. It's not the end point." Wen Sang on why Genspark's recorder turns a meeting into the proposal, the pricing model and the follow-ups instead of stopping at the summary.
chainofthought.show/podcast/73-...
chainofthought.show/podcast/73-...