CASE STUDY · PUBLIC
operant
The Rust agent runtime under the flagship kernel: typed tools, memory, scheduling, and a plugin surface in one consolidated workspace, with the portfolio's strongest executed gate.
Status ruling active development
Evidence for Essay 6, Ideal agentic harness architecture. The spec names what that argument has to hit and the figures it will cite.
Problem
Agent runtimes drift into bloat and mystery: untyped tool surfaces, memory bolted on, no kernel discipline. operant is the intended realization of the harness as a typed, layered substrate.
Build
- Seven workspace crates: core, os, config, harness, plugins, CLI, and the API surface.
- Memory infrastructure whose MCP server is the leanest in the fleet (7 tools, about 121 tokens per tool).
- Web access integration and cron automations with hooks for sub-agents.
Outcome
- The gate ran 7,168 tests with zero failures across 72 binaries at commit ca33d4c, measured on a clean export of that commit. The --no-fail-fast flag is part of the method: a plain run stops at the first failing binary and reports a fraction of the suite.
- Honest status, per the operator's ruling: the runtime is in active development. The green gate is not a production-readiness claim, and the study does not dress it up as one. The October consolidation that reshaped the workspace is recorded in reconciliations R9, R10 and R21.
Provenance
| Gate, executed | 7,168 passed, 0 failed, 24 ignored | cargo test --workspace --no-fail-fast at ca33d4c, 2026-10-20 |
|---|---|---|
| Crates | 7 | workspace manifest, 2026-10-26 |
| LOC | 390,981 under crates/ (+4,381 vendored) | recorded line-count pipeline at 5a116305 |
| Test attributes, static | 4,959 (upper bound) | static attribute count at 5a116305; the executed figure is the row above |
Reproduce
Apache-2.0. Clone the repo and run `cargo test --workspace --no-fail-fast` at the pinned commit: the runner output re-derives the headline. The static count is an upper bound and is labelled as one.
https://github.com/ishan-parihar/operant
Published documents
- AD-006 · Bodies are fenced from the supervisor A body cannot read the supervisor environment, cannot inherit chair authority, and dies with its slot - the isolation boundary as a decision.
- AD-007 · An unfenced run is refused A run whose policy.json is missing fails to start rather than running with no seat authority.
- AD-023 · Plugins cannot replace a built-in Plugin enable state is a config list, a plugin tool cannot take a built-in name, and its HTTP cannot reach an internal address.
- AD-035 · One libclang resolver, and no dangling link Three build scripts each resolved libclang differently and one fell through to a directory that does not exist; a shared resolver now owns the choice and rejects a removed symlink target.
- AD-040 · The dashboard token only goes to its holder The session token was served on an unauthenticated route, so the bearer auth was theatre; it is now exchanged for an HttpOnly cookie and the Host header is validated.
- The operant mechanism library the mechanism + intention behind each element of the harness
- Organization — the Holon runtime, seats, and work graph operant's organization surface: seat policy, budgets, authority, kanban, the curator, and cron.
- Kernel — the persistent Python kernel + continual harness the persistent kernel's mechanism
- The agentic core — the ReAct loop the mechanism of operant's agent loop
- The tool system the mechanism of agent-callable tools
- Context management — eviction, decay, and the compaction ladder how operant keeps a long conversation inside the context window —
- Gateway — the channel surface the multi-platform messaging gateway
- MCP — both sides of the protocol operant's Model Context Protocol surfaces
- Plugins — WASM extensions the WebAssembly plugin system
- Sub-agents — delegation, swarm, async the mechanisms by which one agent spawns others
- Browser — the sourcehound engine operant's browser automation mechanism
- Web tools — sourcehound intelligence the web intelligence tools
- TUI — the terminal interface the terminal UI's architecture
- Vision — the auxiliary-LLM contract image and video understanding in operant
- Voice — TTS, STT, and the voice-mode state machine operant's speech surfaces
- Verification — the harness that proves work operant's verification harness
- Autonomous mode — the self-directed loop autonomous coding mode
- Background review — the post-turn learner the background agent that forks after a turn to decide whether a
From the documentation
A 7-crate Rust workspace is the whole reduction from the 18-crate predecessor state, and the operational story is the consolidation itself: fewer crates, one workspace gate, one release train.
cargo workspace and release history, described-not-linked
The runtime gate is a workspace test plus a strict lint that reports zero errors and zero warnings, so a green run is a build claim rather than a suite claim.
repository gate scripts, described-not-linked
The consolidation kept the agentic harness surface intact while the org layer retired to the kernel, which is why operant is the runtime and holon is the kernel rather than two kernels.
disposition records, described-not-linked
Static counts are labelled as upper bounds and runtime figures are labelled with their commit, so the published numbers state which class they are instead of blending the two.
measurement captions, described-not-linked