Thinking out loud on hard problems
- The Multi-Agent Coordination Tax: Mesh vs. Mediator
Agent-to-agent messaging grows quadratically with team size. Why mesh multi-agent systems get expensive, and the mediator pattern microservices already proved.
- SLI/SLO Observability, Built By Hand Before Any Vendor Touches It
A vendor-neutral SLI/SLO pipeline from scratch with OTel, Prometheus, Sloth, Alertmanager and Grafana: three real bugs and a working burn-rate alert.
- MCP Is Giving Up Statefulness — and That's the Point
MCP's July 28 revision removes session state, sampling, roots and logging. Why a protocol built for laptops had to go stateless to run in data centers.
- Self-Hosting Chatterbox on Apple Silicon: The Build Log, and the Bigger Bet Underneath It
A build log for self-hosting Resemble AI's Chatterbox voice-cloning model on Apple Silicon: the judgment calls the docs skip, and why the pattern matters.
- The Agent That Forgets Everything: Cross-Session Memory Without Breaking the Audit Trail
Part 6: what happens when the context window closes. Cross-session memory for a compliance agent with Mem0 and pgvector, and why it's the hardest part.
- 530,702 Tokens in the Index. 560 Reached the Model. All 12 Answers Correct.
Part 5 runs the numbers: a top_k sweep from 20 down to 1 with FlashRank and BGE rerankers. 530,702 tokens indexed, 560 reached the model, 12/12 correct.
- Why Did the Wrong Tokens Arrive? Reranking vs. PageIndex
Why did the wrong tokens reach the model at all? Part 4 compares reranking and PageIndex, two retrieval architectures that answer it in opposite ways.
- 60–95% vs. 9.2%: What I Actually Measured With Headroom
Headroom claims 60–95% token reduction. I measured 9.2% on the config I'd ship to a regulated client and 25.6% on one I wouldn't. Here's why that matters.
- 97% Fewer Tokens, Zero Model Changes: RTK at the Harness Boundary
RTK cuts an agent harness's token bill by filtering tool output at the boundary, with no change to the model, prompt or pipeline. It covers far more than git.
- Your Agent Doesn't Need a Better Model. It Needs a Harness.
Most agent failures aren't capability problems — they're enforcement problems. A controlled six-rung experiment shows what actually moves the needle.
29 posts