MCP cut the session cord
MCP's July 28 rewrite removes protocol sessions and makes each request self-describing. That's a real deployment win. It also hands retries, state ownership, compatibility, and cleanup back to the application.
#DIY Agent
BLOG
Notes, half-formed arguments, build logs, and the occasional coherent thought. Basically where the rambling goes once it seems useful enough to leave in public.
JUL20263 entries
The Project Is Still the Curriculum
Agents can remove implementation toil without removing learning. The project changes shape: prediction, evidence, investigation, and the consequential seams you decide to own.
#DIY Agent#Traceability
Token laundering and the benchmark that escaped
An agent split an auth token into pieces so a scanner could not see it. Every fragment passed. The sequence rebuilt the credential anyway.
#DIY Agent#Traceability
MAY202612 entries
Recovery and the black box recorder
The boring difference between a demo and a working harness is whether you can recover after the model gets lost. Sessions are the black box; compaction is the part that lets you keep flying.
#DIY Agent#Pi Harness#Traceability
Subagents only when they pay rent
Multi-agent systems are not automatically deeper. Most of the time they are just more places for the same confusion to hide. Branch when the evidence will resolve it; not because the diagram looks important.
#DIY Agent#Pi Harness
Hooks are where taste enters the loop
Hooks are not for making the agent clever. They are for making the boring rules hard to forget. Start with warnings; turn the cage on only when you trust it.
#DIY Agent#Pi Harness
Verification is the product
An agent that edits without proving the edit is just autocomplete with a shell account. Wiring a verification gate into Pi so the harness, not the human, checks the model's homework.
#DIY Agent#Pi Harness#Traceability
The filesystem beats the transcript
The transcript is a terrible database. The filesystem is already a pretty good one. Pass references instead of paste walls and the harness stops being a transmission bottleneck.
#DIY Agent#Pi Harness
Project context is a tool, not a vibe
AGENTS.md is not decoration. It is the repo teaching the harness what local truth looks like, so you stop fighting the prompt every time the agent forgets the rules.
#DIY Agent#Pi Harness
Tools are the agent
ReAct sounds fancy until you watch it become read file, edit file, run test, read error, edit again. Tools are not what the agent uses. Tools are what the agent is.
#DIY Agent#Pi Harness
The model has no hands
A model can talk about a bug. A harness gives it the boring parts that make work real: repo files, edits, tests, and feedback. The first primitive in the Pi coding-agent harness series: hands.
#DIY Agent#Pi Harness
Spec-driven agents need valves, not cages
Strict specs are great when the target is known. They get expensive when the spec quietly prevents the agent from finding the target.
#DIY Agent
Git tells you what. The trace tells you why.
Git can show what changed. It cannot tell you which prompt, model, tool failure, or review decision made the line exist. That missing join is the point.
#DIY Agent#Traceability
Wiring multi-agent into relay-ide
relay-ide was a one-agent terminal. Multi-agent meant rethinking what "session" means at the PTY layer — and admitting the IDE was the wrong primitive to grow from.
#DIY Agent#Relay IDE
When orchestration becomes the brittle layer
On three procedural domains (travel, Zoom support, insurance), serializing the full workflow into the system prompt beat LangGraph orchestration on success rate AND consistency. Higher tokens per conversation, fewer LLM calls, better outcomes. Scope-limited to procedural workflows but the direction is clear.
APR202614 entries
Treating YouTube like article intake
Transcribed this for GBrain as a test of the "article-intake" pipeline against a long-form video. If you can outline + extract takeaways from a 90-minute chat, your intake step probably works for everything else.
The write path is the architecture
The dangerous moment in an agent memory system is not retrieval. It is write-back: when an observation becomes durable enough for the next agent to treat it as reality.
#Memory#DIY Agent#Traceability
Same memory, different window
The mistake is loading everything because everything might matter. Explore, climb, and summit need different slices because their failure modes differ.
#Memory#DIY Agent
Coding agents are not product agents
Ethan's framing for why a perfectly capable coding agent still doesn't ship a perfectly good product. The gap is taste and product judgement, not capability — and that's not the gap the harness can close for you.
9B Qwen beats GPT-4.5 — when the scaffold is right
Itay's whitepaper for little-coder. Same Qwen3.5-9B model: 19% on Aider Polyglot inside Aider's scaffold, 45% inside little-coder's. That second number beats GPT-4.5-preview (44.9%) and GPT-OSS-120B (41.8%) — on a consumer laptop, no cloud inference.
Tool traces are memory too
The most useful thing a coding agent remembers is often not a fact. It is the trail: files read, tests failed, patches tried, assumptions broken.
#Memory#DIY Agent#Traceability
Preference and truth age differently
Knowing the user is not the same kind of knowledge as knowing the repo. Relationship memory deserves its own layer because it is scoped, revisable, and easy to abuse.
#Memory#DIY Agent
A vector hit is not a fact
Semantic memory is useful because it surfaces prior context. It becomes dangerous when retrieved context starts acting like reviewed truth.
#Memory#DIY Agent
Canon is not memory
The most trustworthy memory in my stack is not clever. It is markdown in a repo: inspectable, reviewable, and boring enough to survive tool churn.
#Memory#DIY Agent
hindsight — learn, not just remember
Memory service that explicitly separates world facts, observations, experience facts, and mental models — then runs TEMPR retrieval (semantic + BM25 + graph + temporal in parallel) and a CARA reflection step on top. Local Docker stack, Python/Node SDKs, MIT.
Memory is also about forgetting
Portable memory layer for CLI/coding agents. SQLite backbone with markdown + YAML mirrors that you can actually inspect with `cat`. The interesting primitives aren't the storage — they're the lifecycle: half-life decay, retrieval strengthening, sleep consolidation, conflict tracking.
Agent memory isn't one thing
Every week a new "agent memory" project ships. Most of them claim the same job. Spent a week putting five of them next to each other and the actual finding is that "memory" is being used as a label for five different jobs that don't share a contract.
#Memory#DIY Agent
pi-mono — minimal agent substrate by Mario Zechner
The monorepo behind pi — minimal AI agent toolkit with a unified multi-provider LLM API and an extension model that lets you reshape the scaffold without forking the runtime. Mario also built libGDX years ago, which is its own kind of recommendation.
Context Is the Constraint
Capability stopped being the bottleneck a while ago. The hard part now is routing the right context to the right task without blowing up the window — and that's a job for the harness, not the model.
#DIY Agent
