Video summary

What I Learned Building My Own Agentic Harness

Main summary

Key takeaways

Technology

Summary

Josh Wulf describes building NanoCoder, a small coding-agent harness. He started with the basic tools Geoffrey Huntley recommended—list files, read files, edit or write files, and run Bash—then added capabilities as needed.

The central lesson is that agents can use additional tools when those capabilities are exposed through the harness. Tools that operate outside the conversation context can also help agents work over longer periods.

NanoCoder’s core features

  • Session resumption and configuration: Sessions can be resumed after quitting. NanoCoder can load repository instructions, such as agents.md, and skills from a skills directory. It also supports multiple model providers, model selection, command completion, and retrieving context-window sizes from provider APIs.
  • File operations: The read tool accepts paths relative to the working directory or absolute paths, returns line-numbered text, and can read large files in chunks using offsets and limits. Editing uses find-and-replace on a file that has already been read. Writing creates a file or replaces its contents, while Bash provides access to other shell operations.
  • Smart context compaction: At 80% context usage, NanoCoder asks a model to summarize the current context. The summary replaces the active context, but the original messages remain on disk with line numbers. The summary can point back to relevant parts of that history, allowing the agent to retrieve details omitted from the lossy summary. Wulf reports that, in his analysis, agents—particularly Opus—looked back at the original history about 25% of the time. He says evaluations found this useful for recovering details such as previously successful commands.
  • Plans: The agent can create and update a to-do list, mark items complete, and inspect the plan. Plans keep the agent’s goal and outstanding work available across context compaction, supporting longer-running tasks.
  • Memories (which the agent calls notes): Memories can be scoped to a repository, directory, or globally. Each includes a fact, evidence, anchors, and a type, such as a hint or rule. A small index of recently used memories is added at session start, and a matching tool call can trigger a reminder. Memories can be searched or deleted and are backed up to a repository. Wulf found that global memories unexpectedly gave parallel agent sessions shared situational awareness.
  • Timers: An agent can set a timer with a note describing what it should do when the timer fires. The reminder is injected into the agent’s prompt at the scheduled time, allowing it to return to work without sitting in a polling or sleep loop.

Together, plans, memories, timers, and access to compacted history help maintain work across longer spans of time.

Wulf also describes a practical usability fix: copying wrapped terminal text once introduced a hard line break into a command, contributing to the accidental deletion of a directory. He changed copying so it no longer inserts hard wraps and backed up agent memories to avoid losing accumulated knowledge.

Nano Workforce: durable agent orchestration

NanoCoder can use Nano Workforce, an orchestration engine with a business-logic layer for tasks such as implementing issues, developing features, converging pull requests, addressing review comments, merging, and sequencing dependent work. Its processes can wait for conditions such as a PR being opened or merged, a release occurring, or an issue closing.

The engine supports a fleet of NanoCoder instances working on other machines. A durable workflow can continue managing work even if the initiating agent session ends or the human operator is unavailable. Wulf demonstrates delivery graphs that represent dependencies and sequence work; agents can generate, preview, and dispatch those graphs. He says an agent managed work on his behalf while he was on vacation.

D8CTL: an agent-discoverable CLI

Wulf presents D8CTL, a CLI written in D that can expose APIs described by OpenAPI specifications. It includes the Nano BPM engine API and can also work with the Camunda 8 engine.

Rather than hand-coding every API command, D8CTL can examine an OpenAPI spec and generate an explorable CLI surface, including help, command completions, resource actions, argument types, and response fields. Agents can discover commands by requesting help and drilling down through the CLI, then call the CLI through Bash. D8CTL can also adapt the commands it presents based on the server’s negotiated API version.

Wulf says generated coverage avoids gaps and staleness associated with manually maintained command surfaces.

Rust-to-D rewrite

NanoCoder was initially written in Rust, then rewritten in D because Wulf found Rust builds and resource use costly when many agents were working in parallel. He reports the following results:

Measure Rust D Clean release build 63 seconds 12.1 seconds Rebuild after editing one file 42 seconds 11 seconds Binary size 6 MB 4.8 MB Task wall time 70 milliseconds 70 milliseconds

Rust had 288 crates of dependencies, compared with D’s Phobos and system libcurl.

Wulf says the D version opts into its safety features, but D does not offer all of Rust’s checks, including equivalent Send/Sync thread-safety checking. He frames the choice as an engineering trade-off between safety and correctness confidence on one hand, and build speed and iteration velocity on the other. He recommends assessing the trade-off by measuring defects rather than treating language choice as a matter of doctrine.

Reviews, guides, and demonstrations

  • Demonstration: Wulf walks through NanoCoder’s tools, context handling, plans, memories, and timers; shows Nano Workforce workflows and D8CTL command discovery; and compares Rust and D build results.
  • Evaluation: He reports evaluations of smart context compaction, including the finding that agents sometimes retrieve the original history to recover details.
  • Review or tutorial guide: No formal product review or step-by-step written guide is presented. Wulf points viewers to nanobpm.io for information about NanoCoder and the Nano BPM engine. He notes that the NanoCoder page needs updating to reflect the rewrite in D.

Main speaker and sources

  • Josh Wulf — Presenter and creator of NanoCoder, Nano Workforce, and D8CTL; provides the demonstrations, measurements, and analysis.
  • Geoffrey Huntley — Referenced as the source of the initial recommendation to build an agent harness around basic file and shell tools.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video