Reading
What I'm reading.
Notes and summaries of agent-engineering work by others — not my own writing. The pieces that shaped how I think about agents, grouped by theme.
Orchestration
The canonical starting point — workflow vs agent, and five composable orchestration patterns.
When to reach for an agent, single→multi-agent evolution, and guardrails at every stage.
Orchestrator-worker in production; ~15× the tokens, so reserve it for high-value work.
16 parallel Claudes over ~2,000 sessions wrote a 100k-line C compiler that builds the Linux kernel — agent teams as a supervision pattern.
Context
Context is a finite attention budget; the job is curating what goes in.
KV-cache hit rate, the filesystem as external memory, and keeping failures in context.
Tools
Evaluation-driven tool design; fewer, higher-leverage tools beat many overlapping ones.
Tool search, programmatic tool calling, and tool-use examples — big token and accuracy wins.
A pause-and-reason step for tool-heavy, policy-dense decision chains.
Responses API + open-source Agents SDK; the Agents and Handoffs primitives.
Skills
SKILL.md folders + progressive disclosure, shipped as an open standard.
Adopting the Skills standard; Skills and MCP are complementary, not competing.
Curate a per-project skill library; turn team expertise into a reusable asset.
Harness
Init-agent + coding-agent, a JSON feature list, and context recovery across windows.
Persistent project workspaces so setup becomes reusable, not repeated.
Propose-and-approve self-updating instructions and skills.
Planner–generator–evaluator: splitting generation from evaluation beats self-review — then a 1M-token window killed the sprint structure entirely.
Evals
Automated evals pre-launch, monitoring post-launch, A/B once there is traffic; start from 20–50 tasks drawn from real failures.
Safety
Classifiers instead of permission prompts — an injection probe on inputs, a transcript classifier on actions, with honest FP/FN numbers.
Supervise what the agent can do, not what it does — sandboxes, VMs, egress control; users approve 93% of prompts and stop reading them.
Platform
Agent Builder, Connector Registry, ChatKit, and expanded evals.
From pure orchestrator to a full execution environment — native sandbox, harness, memory.
A hosted service for long-horizon agent work — interfaces stay stable while the harness underneath keeps changing.
Field data on delegation: 80% of requests map to 30+ minutes of human work — the horizon is lengthening fast.