Most agent failures are not failures of intelligence. They are failures of stopping.
We just published a deep dive on EPIC Mode — Episodic Policy and Intention Control — a proposed execution mode for archon-level agent orchestration. The full article is on the CatGame Research Frontier:
https://frontier.catgameresearch.net/epic-mode.html
Why agents get stuck
Recent work on the reasoning-action dilemma shows that agents often overthink. Cuadrón et al. found that, on SWE-bench Verified, simply choosing the less-overthinking solution improved performance by almost 30% while cutting compute cost by 43%. Pandian et al. showed that a fast/slow “Snap-Think” mechanism raised solve rates on NYT Connections puzzles from 72% to 98% with GPT-4o. More reasoning is not always better reasoning.
What EPIC does
EPIC proposes a small set of control layers:
- Engram — long-term memory, including a dream graph of cross-session patterns.
- Policy — when to parallelize, sequence, escalate, or stop.
- Intention — a shared goal DAG that every sub-agent reads.
- Control — an observe-delegate-execute-compress-dream-integrate loop.
The archon orchestrator does not do the sub-agents’ work. It holds the shared intention, runs a satisficing gate, and intervenes when a sub-agent overthinks.
Anti-paralysis devices
The core idea is bounded rationality — Herbert Simon’s term for searching until an option is “good enough” rather than optimizing forever. In EPIC, the archon sets token, time, and confidence budgets per subtask. A composite overthink score triggers a ladder of interventions: switch to System-1, refactor the task, spike-and-delegate, or escalate to the dream graph.
Shared memory as infrastructure
EPIC treats shared memory as first-class policy, not plumbing. It draws on governed shared memory and the Mesh Memory Protocol to prevent failure modes like unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. Every fact carries source, scope, timestamp, and confidence.
The full blueprint
The post includes research-backed citations, an inline SVG control-loop diagram, a concrete runtime table, and a blueprint for implementing EPIC in an existing agent protocol. It also includes an honest “gaps and risks” section — EPIC is a design proposal, not a magic cure.
If you are building multi-agent systems and running into analysis paralysis, the full post has the architecture, the numbers, and the implementation sketch:
답글 남기기