AI Builder Notes - April 2026

Another month of liked tweets distilled into something readable. April was heavy on the “what goes around the model” theme - harnesses, managed infrastructure, memory, and where the actual money is.

Six things I’d tell someone starting out

The harness is the product, not the model. When someone says “we’re building an AI agent,” what they usually mean is they’re calling an API. A real agent system has context management, tools, prompts, skills, memory, caching, compaction, permissions, evals, and a way to judge whether the output is any good. “Use the best model” is not architecture. “Give the model this sandbox, these tools, these traces, this memory write path, and this test surface” - that’s architecture.

The OpenAI Agents SDK [1], Meta-Harness [2], browser-harness [3], Agent CI [4], and Passmark [5] all made this point from different angles. The model sits inside a larger machine. That machine decides everything that matters: cost, reliability, permission boundaries, memory, eval quality, and final output.
Copy the evals first. Everything else in an agent system can be rebuilt. Traces become training material for the harness. Generated code is a black-box artifact you can’t reason about. But specs and tests - those become the objective function. The systems worth studying weren’t “agent demos.” They were loops built around traces, scores, assertions, and replay. [1] [2] [5]
Managed agents are becoming real work surfaces. Claude Managed Agents had the cleanest product shape I’ve seen: brain/hands/session split, credentials kept outside the sandbox, durable event log, disposable Linux containers, OpenTelemetry baked in, $0.08 per session-hour, and a reported 60% drop in median time-to-first-byte. [6]
Persistent agents need persistent state. Codex [7], Claude Code routines [8], Gemini CLI subagents [9], pinned threads, scheduled work, automations, heartbeats - all pointing at the same requirement. The agent needs somewhere to work, somewhere to remember, a way to resume, and a way to ask for permission.
Memory is not storage. A folder of markdown files, a knowledge graph, a vector database - those are starting materials. Memory is the decision about what enters context when, what gets compacted, and what provenance stays attached. [10] [11] [12] [13]
Enterprise AI money lives in the boring 90%. Migrations, broken internal systems, compliance work, support queues, data cleanup, job queues, search, documentation, evals, permissions, human review loops. That’s where AI work turns into revenue. HireCade is the cleanest example: $22M ARR, 95% margins, 5 people, no outside funding, built in 14 months. Mostly annotation services with a 30% service fee and a large candidate database. [14]

On harnesses specifically

You can give two tools access to the same model and get completely different results. The difference is almost always the surrounding infrastructure. Context management, tool definitions, prompt templates, skill libraries, memory systems, caching, compaction strategies, permission layers, and provider behavior - that’s where the variance lives.

OpenAI’s Agents SDK made the parts explicit: agents, handoffs, tracing, long-running agents, sandboxes, memory. [1] Stanford’s Meta-Harness treated harness improvement itself as an optimization loop over code, traces, and scores. [2]

browser-harness reframed Chrome automation as a self-healing CDP harness rather than a brittle browser script. [3] Agent CI turned local GitHub Actions into a loop that agents can execute against. [4] Passmark wrapped Playwright with natural-language regression tests, model assertions, caching, and telemetry. [5]

The common thread: none of these are interesting because they have agents. They’re interesting because they wrap the model in a measurable workbench. Evals are the part I’d steal first.

Managed agents becoming real

Claude Managed Agents drew a clean line between three things: the brain (reasoning), the hands (tool execution), and the session (state and context). Credentials stayed outside the sandbox. Work happened in disposable Linux containers. The event log was durable. OpenTelemetry was built in from the start. [6]

Codex was heading toward something broader - a universal development application. Browser use, computer use, multi-terminal, SSH/devboxes, docs and PDFs, memory, plugins, automations. Chronicle-style memory gave recent screen context without the user having to repeat themselves. [7]

The pattern kept showing up in different clothing. Claude Code routines, Codex automations, Gemini CLI subagents, pinned threads, scheduled work, heartbeats - same idea, different surface. A persistent agent needs persistent state and a real work surface. [8] [9]

Shopify’s AI Toolkit and Cloudflare’s Agent Lee made the enterprise version concrete. Agents were getting write access to products, orders, inventory, SEO settings, images, Workers, R2, DNS, error summaries. The protocol and permission layers mattered far more than any chat UI. [15] [16]

Memory and retrieval

Storage is easy to fake, which is why memory is annoying.

GBrain [10], LLMwiki [11], Rowboat [12], Obsidian AI tools [17], and Hermes memory workflows [13] were all attacking the same problem: how do you turn raw tweets, chats, notes, decisions, and project history into recall that enters context at the right moment?

The useful observations were simple. Markdown stays human-readable. Compaction is the actual memory write path. Graph edges matter for people and claims, not just documents. File search fails when the agent doesn’t know it should search. And proactive injection - putting the right memory into context before the agent asks - is the genuinely hard part.

A giant archive is not memory. Memory is deciding what enters context when.

Workflow packaging

The meta-tool this month was SKILL.md: a solved path compressed into instructions, scripts, and boundaries. The same pattern kept appearing everywhere. Do the work once. Keep the trace. Remove wasted steps. Save the workflow. Let the agent start from the compressed path next time.

The workflows worth stealing were small and specific:

Database performance loop: seed repro data, optimize query, try ten indexes, measure impact
Spec-first agent work: review SPEC.md boundaries, let agents work behind them, review contracts instead of every line
Skillification loop: do the work once, turn the solved path into a reusable skill
Implementation notes in HTML: keep an implementation-notes.html file while working - decisions, tradeoffs, gaps stay readable
Deterministic first: a cron job plus one LLM API call replaces most “agents.” Recurring workflows as code, not token burn
Browser loop: agent sees state, acts in Chrome, verifies result

Where the money is

HireCade: $22M ARR, 95% margins, 5 people, no funding, 14 months. Mostly annotation services with a 30% service fee, a large candidate database, and a long enterprise sales cycle. [14]

AI labs and AI-heavy companies need verification, review, annotation, migration, maintenance, and domain-specific human loops. “AI adoption consulting at $200+/hr” is real when you’re actually installing workflows, not just giving talks.

Enterprise AI money is the ugly 90%: migrations, broken internal systems, compliance, support, data cleanup, job queues, search, documentation, evals, permissions. That’s the work.

Most AI adoption stalls on missing work graphs - connected tools, permissions, memory, evals, workflow defaults, review gates, and someone willing to redesign how the work actually flows.

Ship, share, repeat still beats separate marketing. Code plus content plus product storytelling is a distribution system.

Tools worth opening

Hermes Agent - open agent stack with skills, tools, memory, phone access, and cost visibility
GBrain - personal total-recall layer around OpenClaw / Hermes-style workflows
Cabinet - open-source startup OS with agents, schedules, knowledge base, browser terminal, local-first
browser-harness - self-healing browser harness for Claude Code / Codex-style work
Pi - worth tracking for browser-adjacent agent workflows and logged-in work surfaces
HyperFrames - agent-native HTML to MP4
Firecrawl web-agent and Firecrawl Parse - web/PDF ingestion for agents
Agent CI - local GitHub Actions for agents
awesome-design-md - brand/design systems as markdown for agents
Syncthing - boring peer-to-peer file sync, relevant because durable local files beat trapped SaaS databases