1. Where It Started: Memory and Context
I started with Claude Code around April 2025. The first real step was recognising that Claude's native memory was essentially useless. The workaround was using markdown files as persistent memory stores, editable both through Claude and tools like Cursor. That opened the door to storing not just session notes but also instructions, roles, and agent skills — anything that would otherwise be forgotten across context resets.
But the fundamental problem remained: at some point the context window fills, the model gets amnesia, and starts behaving destructively. Cursor handled this somewhat better at the time. Gemini had an edge due to its larger context window (already at 1M tokens), though at a cost. Neither was a real solution.
2. The Core Principle
Taking a step back from tooling led to the central insight the whole framework is built on:
The better the prompt, the better the output. The better the instruction — and the context around it — the higher the likelihood of a good result.
This is no different from how you'd brief a human. Context, clarity, traceability, constraints — all of it shapes the quality of what gets executed. The question became: how do you systematically generate and maintain that context?
3. The Agentic Engineering Framework
To produce good, consistent context, you need to capture:
- What has been done before — every instruction, tool call, output, error, pivot, and decision
- What the goals and architecture look like — what was decided and why
- What is connected to what — if you change function X, what does that break elsewhere?
This last point introduced the concept of the blast radius — borrowed from physical and industrial engineering. It describes the potential impact zone of any given change.
Context Fabric
Captures the full history of work: what was done, what failed, what changed, what was decided. When an agent starts a new task, it can look back at relevant prior context rather than starting blind.
Component Fabric
Provides structural awareness — understanding how components relate to each other so that an action's blast radius can be assessed before it's taken.
The Prime Directive
Nothing gets done without a task. Every action must be linked to a task. This enforces traceability and prevents autonomous drift.
Enforcement is the hard part. Git hooks work sometimes. Claude Code doesn't reliably respect the constraint — partly due to the stochastic nature of LLMs, partly due to permissive execution environments. If broad tool permissions are granted, there's nothing structurally preventing the model from bypassing the rule.
4. TermLink
The challenge: how to coordinate multiple agents in a reliable, deterministic way.
The idea behind TermLink is that if terminal sessions are initialised in a known state, you can inject into them — essentially simulating a USB keyboard over the terminal link, sending ASCII sequences directly to the session.
In practice this works well. The weak point is that Claude Code sometimes falls back to calling claude -p through PTY rather than opening a terminal and running it properly. That loses the interactive feedback loop — the back-and-forth that makes the coordination meaningful.
TermLink is also now using a network socket interface. This opens up communication across machines — and with it, the possibility of real orchestration: routing tasks to different agents, mixing providers, and matching the right model to the right type of work reliably.
5. Proof Is in the Pudding
The proof is in the eating of the pudding. I'm using the framework to build real things I can use, and that's where you find out if it actually works.
Open-Claw ingestion: Took the open-claw codebase, ingested it through the context fabric, exposed it for browsing and querying. Used it to extract improvement ideas for the Agentic Engineering Framework itself. The model identified enhancements, formatted them against the standard task structure, and dispatched them to the TermLink agent, which pulled from the knowledge repository and started working autonomously. It worked.
Email archiver: Started as a utility to consolidate ~70K emails across Hotmail and Google domain accounts into a single searchable archive (useful for things like digging up tax receipts). Evolved into a fuller email client with AI capabilities — translation, generation, support for both local and remote models. Still in progress. A rough first release has been pushed to GitHub. The focus has shifted toward a more controlled, personal-assistant-style interface rather than trying to match full open-source alternatives.
If you want to test-drive the OpenClaw Fabric Explorer or the AI Email Personal Assistant, drop a comment: EXPLORER. You find the other repo's here:
The Engineering Framework
Termlink
Top comments (0)