
EngineeringField note
Making Phantom’s engineering loop faster
AI-generated illustration

EngineeringField note
AI-generated illustration
Two small changes reduced repeated work in Phantom. The useful part is how we measured them, preserved the checks and kept a local result in perspective.
Waiting is easy to feel and surprisingly hard to measure. A tool call pauses, a test suite takes another minute, and the temptation is to optimize whatever is closest. In Phantom, our workbench for coding agents, we started by asking where the time actually went. Two changes came out of that investigation: start independent tool requests together and stop repeating a check the same request had already completed.
We began with a completed release-check run: 39 minutes and 4 seconds, with all 15 jobs successful. In its longest Mac general lane, fixtures accounted for about 90% of reported test wall time. Those fixtures perform real preparation, delivery and recovery. That historical measurement told us where to investigate; it is not a before-and-after result for the upcoming release.
The first change concerns tool discovery through the Model Context Protocol, or MCP. Independent discovery requests now start together. Results still merge in the original registry order, so a faster server does not rearrange tools or take over a native tool name. Tests hold both responses open to prove that the requests start independently, then finish them in reverse order. They also check failed peers and fresh discovery on the next request.
The second change removes a repeated read. Phantom builds a view of delivery status from the underlying records; engineers call that view a projection. Once a request has verified it, the same request can reuse it instead of immediately rebuilding it. The result is not cached across requests. Every new read still checks current registry, campaign, recovery and artifact state. In four alternating controlled comparisons on the same prepared store, redundant post-recovery projection reads fell from one to zero. The removed projection took a median of about 97 milliseconds under those conditions.
The real setup and mission acceptance checks also passed: three cases, including an owner restart, a second scope and Stop. These results support the change’s behavior under the tested conditions. We have not measured a matched improvement in total mission or release time, and independent tool requests still depend on the servers they contact.
Both improvements were included in the Phantom 3.24.4 candidate reviewed for this October 7 field note. At that time, release qualification was still in progress and the published and installed release was 3.24.3. The linked research appendix retains those measurements, source revisions, and a candidate interface preview. Later release notes are separate evidence; they do not turn these controlled measurements into a total release-time result.
Profile the whole path before optimizing a convenient small function. Keep queue delay, compilation, fixtures and packaging separate. Reuse verified work within the request where its inputs remain valid; a fresh request must revalidate. Test reverse completion, partial failure, source drift and a second read so faster feedback does not turn into stale results.
Track useful work completed alongside latency. A shorter wait for tools and a shorter release pipeline are different outcomes. Ask which source revision and conditions produced a number, and keep the tests that establish correct delivery, restart and Stop. Candidate previews should help you understand the product without implying that every connected provider is active.
Explore the Phantom engineering workbench at phm.dev and Phantom Secrets at phm.dev/secrets. Open the research appendix below for the October 7 charts, downloadable data, and historical candidate preview. The project links below point to the current products; the appendix preserves the conditions of the original measurements.
Ashlr tools connected to this story. Review their source and setup requirements.
Work with eligible coding-agent accounts and local models in one workbench; review fleet proposals under signed authority.
Source material behind this field note. Our practical implications are Ashlr’s interpretation.
Clarified the October 7 release status as historical and updated the workbench and Secrets destinations. Original measurements and source revisions are unchanged.
Field notes explain an engineering or business practice. They are not reports of a new event. Cover artwork is illustrative.


Ashlr release
Phantom published releases 3.30.67 and 3.30.66. Lexicon and Locus updated documentation on test selection and adapter guidance.
EngineeringThe briefing
AI-generated illustration
The AI evaluation nonprofit added tasks that require research taste and judgment to its time-horizon evaluation. Frontier models still lag humans on these tasks even while they continue to double the length of verifiable tasks they can complete every four months.


Ashlr release
Delegate supported API work through scoped placeholders and a local proxy, keeping real keys out of agent context.
Several Phantom pull requests updated the unified inbox to show legacy bus messages and inbox triage counts. Other changes added fleet yield, landed, and pace tools plus documentation updates across projects.