Skip to content
View AllStreets's full-sized avatar

Block or report AllStreets

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
AllStreets/README.md

Connor Evans

I build control planes for untrusted autonomous agents.

capability gating  ·  kill switches  ·  audit ledgers  ·  graduated autonomy


tests scout lines

Chicago, IL  ·  connorevans29@gmail.com



The through-line I didn't plan

I built four of these over five months without intending them as a set. They converged on the same architecture anyway, in three different languages, and I only noticed once I put them side by side.

                          ┌──────────────────────────────────────────┐
   an agent wants to      │  1  declare it first ....... manifest    │
   do something           │  2  gate it ................ arbiter     │
          │               │  3  bound the spend ........ budget      │
          └──────────────▶│  4  stop it mid-flight ..... revoke      │
                          │  5  write down what happened ledger      │
                          │  6  earn the next rung ..... trust       │
                          └──────────────────────────────────────────┘
                                             │
             ┌───────────────┬───────────────┼───────────────┐
             ▼               ▼               ▼               ▼
           LOOM           ONEXUS            WFR            SMADP
        Rust · TS         Python         TypeScript    Python · Docker
LOOM ONEXUS WFR SMADP
declare first organ manifest Aegis manifest tool policy agent profile
gate sandbox validation check_capability trust gate tripwire engine
revoke live per-organ revoke() kill switch + halt stale on drift
bound spend per-organ budget trust tiers cost ledger judge triage
record it git timeline Chronicle route ledger chronicle log
earn autonomy consent cards suggest → autonomous trust sentinel evidence ladder

Every one is the same bet: an agent should have to say what it needs before it runs, be refused when it asks for more, be stoppable mid-flight, and leave a record you can read afterwards. I find that more interesting than making agents capable, because capability is the part that already works.


Selected work

Scout
Flexport · internal
30+ countries

Hired as an inbound SDR, then selected to lead workflow automation and Clay-based enrichment for the sales organization. That mandate became Scout: prospect discovery, account intelligence and enrichment pipelines in one internal platform, comparable in scope to LinkedIn Sales Navigator. Now used by Flexport sales employees in more than 30 countries; my team has held 100% of quota throughout.

The only system here whose adoption was decided by other people.

ONEXUS
Python
1,366 tests

A local-first agent runtime built on Aegis, a 786-line capability arbiter. An agent declares filesystem reach, outbound domains and tools in a manifest; every call is checked against it, allowed or denied, and appended to a ledger that is never rewritten. A static test AST-parses every kernel module and proves none but Aegis can open a socket.

Then I red-teamed it, found that revoke() did not actually stop an agent holding a grant, fixed it, and published the six attacks that failed alongside the three that landed.

LOOM
Rust · Tauri · TS
1,315 tests

An offline computer that builds itself. Describe a capability in one sentence and a local model plans it, writes the code and its tests, proves it in a sandbox, repairs its own failures over bounded rounds, and commits to a git timeline. It then asks permission before the new organ may run. Powers are declared, budgeted and revocable while running. No cloud, no subscription, no telemetry.

SMADP
Python · Docker
973 tests

Safe Multi-Agent Deployment Platform. Nobody has systematically studied what happens when the coding agent, the browser agent and the email agent all share your filesystem and OAuth scopes. SMADP publishes that matrix, with a tripwire engine that halts a run mid-flight when a planted secret leaks or egress steps outside the allowlist.

Every verdict carries its evidence rung as a field on the record. Eleven pairs have actually been executed against each other; 2,268 are labeled priors, in the data rather than a footnote.

AUSPEX
JS · Three.js
134 tests · live

A living globe of the planet's perils and its breakthroughs, free and login-free. Autonomous pipelines read USGS, GDACS, NASA FIRMS, Launch Library and global RSS, score each event for severity and confidence, and render it honestly: calm where the world is calm, red reserved for genuine danger. The disaster path needs zero API keys, which is what lets it stay free.

HYPERION
Next.js · Three.js
511 tests

A browser game about running a private orbital agency over a living Earth, with real orbital mechanics, real world events, and a world that keeps evolving while you are offline. A scheduled server tick advances every away player's world, so you return to a digest of what changed without you.

Flexport SDR Console
React · Express
19,219 lines

I wanted to work in freight forwarding, so instead of writing a cover letter I built the tool the job would need: ICP-scored prospects, live AIS vessel positions, live ADS-B aircraft, port-congestion rings, FRED macro data and tariff tables in one console. The working core came together in a 30-hour weekend before my final interviews. I kept building it after, and ran it in production for my own team.

EMBER
vanilla JS
0 external requests

An offline survival console for the day the network isn't there: a local LLM advisor, 38 field guides, offline navigation, a solar power budget, and a Forge that rewrites the console's own code from a plain-language prompt. Genuinely zero external requests; the only thing it ever talks to is localhost.


Numbers here are ones I ran, not ones I estimated. Where a project publishes a large figure,
the README beside it says how much was verified and how much was inferred.

Pinned Loading

  1. ONEXUS ONEXUS Public

    A local-first runtime that runs AI agents behind a capability arbiter. Every tool call is gated against a declared manifest, scored on earned trust, and written to an immutable audit ledger.

    Python 4 1

  2. SMADP SMADP Public

    Safe Multi-Agent Deployment Platform: pairwise safety analysis for autonomous agents that share a filesystem, clipboard, and OAuth scopes. Every verdict is graded on a four-rung evidence ladder, fr…

    Python 3

  3. ONEXUS-Agents ONEXUS-Agents Public

    An evidence-graded catalog of open-source AI agents. Nightly crawl, transparent composite ranking with published weights, and hand-written MCP adapters. Every claim is labeled by how it was verified.

    Python 3

  4. AUSPEX AUSPEX Public

    A living globe of the planet's perils and breakthroughs, built for everyone. Autonomous pipelines read public signals from USGS, GDACS, NASA FIRMS, Launch Library and global RSS, score them for sev…

    JavaScript 1

  5. chi chi Public

    CHI ATLAS: a cinematic mission-control room for one city. A 3D Mapbox HUD over live Chicago data, with real CTA L train positions crawling the map, live scores, lake conditions and ten surfaces rea…

    JavaScript

  6. ember ember Public

    The offline survival console. A self-contained dashboard for when the grid goes down: a local-LLM advisor, 38 field guides, offline navigation, a solar power budget, and a self-modifying Forge. Zer…

    JavaScript