The third submit from Construct Membership, our weekly dwell construct session. The companion GitHub repo will be discovered right here, docs right here and you’ll attempt the agent dwell within the hosted playground.
Your agent framework just isn’t the bottleneck. The bottleneck is that each new exterior system your agent wants to speak to requires one other device wrapper, one other MCP server, one other merchandise in a registry that’s at all times two steps behind the API it wraps.
The traditional mannequin is “agent plus curated device registry.” It scales linearly with the variety of integrations your agent has to do, and the curation is everlasting work. You ship a wrapper. The seller modifications their endpoint. The wrapper drifts. The agent will get caught. You ship one other wrapper.
There’s a sample rising in manufacturing that inverts this strategy. The brand new mannequin is “agent plus safe sandbox plus uncooked API specs.” The instruments aren’t pre-built. The agent writes them on the fly, utilizing the spec as its solely reference, runs them in a boundary you belief, and discards those that transform fallacious. The framework’s job is to not present instruments. The framework’s job is to make tool-authoring secure.
Luke Shulman, Director of Agent Innovation at DataRobot, walked by way of this sample in a latest Construct Membership session.
The viewers picked the issue: CODEOWNERS hygiene within the DataRobot monorepo. Each monorepo of significant age accumulates this type of drift as groups reorganize, get renamed, or get absorbed. Information find yourself annotated with aliases that not level wherever. The cleanup is mechanical, tedious, and a superb first goal for an agent. A member of the platform workforce surfaced it because the construct goal: scan the repo, discover recordsdata owned by groups that not exist, suggest reassignments, open the PR.
Luke constructed it dwell, in an hour, on a modest 35B-parameter mannequin. He didn’t pre-build a single device. The agent wrote them.
This submit is the recipe.
Luke’s NL agent authoring its first device in opposition to the GitHub OpenAPI spec.
Luke calls this sample a Pure language (NL) agent, additionally known as a context-agent.
The framing issues as a result of it inverts the place your engineering effort goes. Within the standard setup, you spend your time on the device registry. In an NL agent, you spend your time on the sandbox.
The agent runs in a Deno-based JavaScript VM with a restricted listing, a restricted community allowlist, and a restricted set of atmosphere variables. JavaScript is the best execution floor for this as a result of your entire browser ecosystem is constructed on working untrusted JavaScript safely. Deno tightens that additional with express permissions for file, community, and atmosphere entry.
The agent will get eight instruments to start out: cat, discover, grep, tree, write, search-and-replace, mkdir, and execute_code. All the pieces else, the agent has to creator itself. The execute_code device is the unlock. The agent reads a markdown system immediate, reads any reference docs in its listing, and begins writing JavaScript capabilities to speak to the exterior system. It tries them. It fixes them after they fail. The capabilities it retains get saved as a instruments.js file within the working listing. The subsequent time the agent hundreds, these instruments are already there.
The asymmetry is favorable. Setup is brief. The infrastructure is small. The agent does the combination work itself in opposition to a spec that’s, by definition, extra full than any wrapper anybody was going to keep up. You should not have to be forward of the agent’s wants. The spec already is.
All the pieces beneath assumes you’ve gotten the NL agent runtime (open-sourced at github.com/kindofluke/context-agent) and a DataRobot account. In case you would fairly see the sample earlier than you construct, the hosted playground runs the agent dwell in your browser in opposition to a pattern data base.
Step 1: Arrange the listing and sandbox
Create a contemporary working listing. That is the one place the agent can learn or write. Configure the Deno sandbox to permit solely .js and .md file varieties inside that listing. Configure the community allowlist to allow solely the domains you need the agent to hit. For this construct, that meant api.github.com and nothing else.
That is the load-bearing step. In case you give an agent the power to write down code with no secure place to run it, you get both a refusal-prone agent or a safety incident. The framework’s worth is the sandbox, not the agent loop.
Step 2: Drop within the OpenAPI spec as context
Obtain the GitHub OpenAPI spec and put it within the agent’s listing as github-openapi.yaml. Don’t write a wrapper. Don’t pre-author instruments. The spec is all of the context the agent wants.
Overview of the agent’s listing and context in the course of the construct.
That is the transfer that will get essentially the most pushback and is crucial. The traditional intuition is to write down a skinny consumer across the API and hand the agent the consumer. The NL sample is handy the agent the spec and let it write its personal skinny consumer, just for the endpoints it truly finally ends up needing. Most wrappers cowl floor space that by no means will get used.
Step 3: Generate a fine-grained token as a prefixed env var
Generate a GitHub fine-grained private entry token scoped to Contents: learn and Pull requests: write for the goal repo. Minimal required scope, nothing extra.
The NL runtime exposes atmosphere variables to the agent solely after they carry a selected prefix (NL_ in Luke’s setup). Something with out the prefix is invisible to the agent. That is the way you cease it from by chance studying credentials it has no enterprise studying. Set NL_GITHUB_TOKEN= and the agent will choose it up. Anything in your shell stays out of attain.
Step 4: Give the agent a small, scoped first process
Within the chat interface, inform the agent what it has entry to and ask it to verify connectivity. The very first thing it should do is creator a probe device, 5 or ten strains of JavaScript that hits the rate-limit endpoint. When that works, give it the actual process: “discover each file within the monorepo owned by @datarobot/cloud-operations within the DR_CODEOWNERS file.”
The agent’s first transfer was to creator a device it named getCodeownersFiles. About twenty strains. It walked the repo through the GitHub API, parsed CODEOWNERS patterns, and returned a listing.
It ran the device, obtained again the listing, after which, with out being requested, wrote a second device to persist the listing as a cloud-ops-inventory.txt file in its listing. The agent found out by itself {that a} file makes a superbly good working reminiscence. The tools-as-emergent-memory sample fell out of the runtime with out anybody designing for it.
Step 5: Add a scope-discipline system immediate
The agent’s default conduct is to do an excessive amount of. Earlier than you let it suggest modifications to the repo, give it a system immediate that pulls a tough line round what it will probably modify:
The CODEOWNERS tips solely replace CODEOWNERS references. Don’t modify actual working code. Solely open PRs. Be secure.
That sentence stops the agent from “helpfully” refactoring code whereas it’s within the file. Scope self-discipline issues greater than functionality when you’re handing an agent write entry to a manufacturing repo. From there, the agent labored by way of the stock file by file, proposing reassignments the place the git historical past made the brand new proprietor apparent and flagging the remaining for human overview. The PR-creation step stayed within the loop with a human reviewer, which is the best reply for a primary cross.
Step 6: Lock the agent into read-only mode
As soon as the agent has authored the instruments that work, flip the runtime into read-only mode. The agent can nonetheless name its current instruments, learn recordsdata, and execute the JavaScript it already wrote. It can’t write new instruments. It can’t rewrite its system immediate. The agent is now an artifact.
The instruments.js and the markdown system immediate are your entire deliverable. Drop them into the DataRobot registry and workshop as a {custom} mannequin, and you’ve got a deployable, ruled agent with a completely seen code floor. The exploration part wants write entry. The manufacturing part doesn’t.
The session was scheduled as a wild card. It was the cleanest inner argument we’ve got had about what an agent platform ought to ship. Three takeaways.
Context is what you ship. An entire, well-structured spec for an exterior API outperforms a hand-rolled device wrapped across the similar API, as a result of the spec preserves optionality the wrapper has already discarded. The implication is uncomfortable for product groups: the highest-leverage factor you’ll be able to ship for the agentic period just isn’t a brand new SDK or a brand new device registry. It’s glorious, copy-as-markdown documentation. The “copy web page as markdown” button some open supply tasks have began including just isn’t a UX flourish. It’s a deliberate concession to the truth that the reader is, more and more, an agent. Make your docs loadable. Publish your OpenAPI specs. Hold them present. The brokers will take it from there.
The sandbox is the unlock, not the loop. Most agent frameworks compete on orchestration, reminiscence, and planning. The factor that decides whether or not the NL sample is shippable is none of these. It’s whether or not you may give the agent a spot to execute code that you just truly belief. Deno’s permission mannequin does a lot of the work right here. Restricted file varieties, restricted directories, restricted community egress, prefixed env vars. None of it’s unique. All of it needs to be in place earlier than the agent loop issues.
Greatest-in-class context beats best-in-class frameworks. The brokers that work in manufacturing aren’t those with essentially the most elaborate orchestration. They’re those with the cleanest, most loadable, most agent-friendly documentation round them. Each minute spent on higher markdown is price ten minutes spent on a extra subtle agent framework. Most groups have the priorities inverted, and the price reveals up as brokers that look spectacular in demos and fall over in deployment.
The implication for the DataRobot platform is direct. The registry and workshop already host {custom} fashions. The pure subsequent step is a custom-model workflow that wants solely a instruments.js and a markdown system immediate, with the NL runtime offering the sandbox beneath. No atmosphere configuration. The agent assembles what it wants from a spec you level it at, runs it inside a boundary your safety workforce has already signed off on, and ships as a frozen artifact when it really works.
Construct Membership runs weekly. Every session takes one volunteer driver, one hour, and an concept voted on by the viewers. The format is intentionally unrehearsed: we construct dwell, the construct breaks dwell, and we repair it dwell. In case you are constructing on DataRobot or interested by enterprise-ready brokers and need inspiration, that is the collection for it.
