Agents: Memory, Token Tricks, Subagents, Local Histories, & SAFR

Claude, please stop trying to memorize random crap. Theahura shows session transcripts usually worsen agent performance and recommends storing distilled artifacts and human-reviewed documentation instead of raw chat logs. Outcome engineers should favor curated artifacts over chat dumps to keep agent state lean, auditable, and easier to gate (Principles 08, 06, 15).

Fable’s judgement. Simon Willison describes delegating coding tasks to lower-power subagents so Fable saves tokens while keeping judgment and review in the main loop. Use this pattern to split execution and oversight responsibilities—cheap subagents for throughput, a main agent for validation and human-aligned decisions (Principle 09).

60% Fable cost cut by converting code to images and having the model OCR it. pxpipe renders bulky context as images and lets the model OCR them to drastically lower token costs. Treat visual-encoded context as a viable token-compression tactic when cost dominates, but measure latency, OCR fidelity, and verification risk (Principles 06, 11).

Show HN: ctx — Search the coding agent history already on your machine. ctx indexes coding-agent session histories into SQLite so agents and developers can privately search past decisions and avoid repeating work. Local searchable histories make context retrieval legible and auditable—build tooling to let agents query curated past artifacts rather than raw transcripts (Principles 06, 13).

MAS partners industry to develop safeguards for AI agents in finance. MAS and industry publish SAFR, a runtime framework enforcing policy-bound execution, real-time validation, and auditability for AI agents in finance. Outcome engineers must design agents with enforcement hooks, telemetry, and auditable decision logs to meet emerging regulatory runtimes (Principles 10, 15, 14).