From 1a7c01c77ef219d650ae9dee1a364a67e5561d56 Mon Sep 17 00:00:00 2001
From: declanatinference <262201734+declanatinference@users.noreply.github.com>
Date: Thu, 21 May 2026 20:18:00 +0000
Subject: [PATCH] Add talk: Declan Johnston

---
 README.md       |  1 +
 _data/talks.yml | 23 +++++++++++++++++++++++
 2 files changed, 24 insertions(+)

diff --git a/README.md b/README.md
index 13c2928..d8b0e23 100644
--- a/README.md
+++ b/README.md
@@ -38,6 +38,7 @@ We invite submissions of **5-minute talks** (including Q&A - nothing fancy - jus
 | [**Dhruv Roongta**](https://github.com/Dhruv317)<br>[slashy.com](https://slashy.com) · [𝕏](https://x.com/DhruvRoongta) | We're building an AI-native email client that helps draft emails in the user's voice, triage emails, and never drop the ball on emails. […](https://github.com/milstan/sisw/issues/18) |
 | [**pranav bedi**](https://github.com/pranavbedi)<br>[moda.dev](https://moda.dev) · [𝕏](https://x.com/pranavbedi) | Agents today learn like 2010-era SaaS: a human watches a dashboard, files a ticket, opens a PR. The user is upstream of the fix by days. That's backwards, the user interaction IS the training signal. Every frustration, abandonment, and silent failure is a labeled example of what the agent shouldn't have done. The job is to extract those labels automatically and pipe them into the next self-edit cycle.  Example: one user typed "i received nothing, how come" at 2am. |
 | [**Tiger Wang**](https://github.com/thenewpotato)<br>[nessielabs.com](https://nessielabs.com) · [𝕏](https://x.com/tigerjwang) | The path to self-improving software is easier (and more agentic) than you might think. Our internal agent platform runs on ~500 lines of Express + SQLite + cron calling `claude -p` — no agent framework, no multi-agent protocol, no message bus. It's been running PR reviews and user-facing email campaign agents across Nessie in production. I'll share why we haven't needed anything fancier, where the simple loop actually breaks, and the case for under-engineering your orchestrator in 2026. |
+| [**Declan Johnston**](https://github.com/declanatinference)<br>[inference.net](https://inference.net) · [🔗](https://github.com/context-labs/HALO) | A general-purpose harness like Claude Code is the wrong tool for trace analysis. […](https://github.com/milstan/sisw/issues/31) |
 <!-- TALKS:END -->
 
 <!--
diff --git a/_data/talks.yml b/_data/talks.yml
index 4b8d31f..f431986 100644
--- a/_data/talks.yml
+++ b/_data/talks.yml
@@ -116,3 +116,26 @@
   submitter_github: thenewpotato
   submitted_at: '2026-05-19T06:01:26Z'
   issue: 22
+- first: Declan
+  last: Johnston
+  domain: inference.net
+  repo: https://github.com/context-labs/HALO
+  twitter: ''
+  summary: 'A general-purpose harness like Claude Code is the wrong tool for trace
+    analysis. This isn’t because the model isn’t smart, but because traces can get
+    extremely long, and you need a specialized toolkit in order to make observations
+    about systemic agentic behavior. HALO (Hierarchical Agent Loop Optimization) is
+    a methodology for building recursively self-improving agent harnesses using RLMs.
+
+
+    The core HALO loop is suprisingly simple:
+
+    1. Collect execution traces from your agent harness. HALO uses OpenTelemetry-compatible
+    tracing.
+
+    2. Feed traces into HALO-RLM engine.
+
+    3.'
+  submitter_github: declanatinference
+  submitted_at: '2026-05-21T20:17:59Z'
+  issue: 31