SPIRIT and Agentic Loops
There are many articles and diagrams on agentic loops - I wouldn't attempt to rehash them here. What I'll start with is picking a rough form, explore mapping our SPIRIT framework to see what fits, and what can be improved.
Agentic Loop
The rough form (and also a rough attempt at linking to our earlier 'TL;DR' term):
- Context-gathering
- The results from the previous loop. Loosely maps to learning.
- Planning
- Self-explanatory; I like to think this is the thinking part.
- Acting
- Part-harness, part-tools. Clearly the doing step.
- Observing
- Reading from the results.
- Repeating
- Onto the next loop.
Thinking and learning swap places here because that is the stronger lean. We can also argue that context-gathering require thinking about what survives from the previous loop and what can be discarded; and planning is an exercise in learning about the tasks we keep and iterate on.
Mapping SPIRIT
Things become interesting when we map SPIRIT onto the five steps:
- Context-gathering
- Substrate: the accumulated storytelling.
- Planning
- Intuition: decision-making, when to clarify.
- Acting
- Rites: Manifesting the patterned learnings and decisions into changes.
- Observing
- Impact: what do the changes tell us of the loop's outcome.
- Repeating
- Tenacity: When to continue, or to pause for a review-pass.
Posterity is practically absent from our loop definition - at best loosely implied from the repetition. Without posterity, agentic loops appear to advance readily - tickets are picked up, code changes are merged in, the outputs (UI, files) get a bit shinier, churning like windmills of the mind.
Posterity can therefore be the make-or-break difference when comparing how a person or a team harness (pun intended) agentic loops.
In a person, posterity is often encoded in their memories, which begets a harder failure when that's removed - through teaming, or plain forgetfulness. The context-gathering is made coherent through their own understanding, not the live code repository.
An agent can be made to iterate faster, and the person only accounts for oneself; when left unchecked this version of the agentic loop can give rise to fastest drifts with fewest observers.
In a team, posterity becomes contested by the people, an underrated benefit. Discussions of this nature are a posterity-augmentation generator, and survived decisions often entail the why instead of blind compliance. Agents to a large extent do that or diverge silently.
A team that leans more into agentic work then run the risk of losing their organic articulated reasoning all the while not noticing, since churn remains constant, if not more. This is only checked by the presence of new joiners, which onboarding becomes an unplanned audit activity.
An ASCII art illustration to sum it all:
+--- Loop ------------------------------------+
| Substrate -->-- Intuition -->-- Tenacity |
| ^ (Rites) v |
| +-----<----- Impact -----<-----+ | (in seconds, minutes)
+---+-----------------------------------------+
| (Posterity - Person)
+---+--- Session -----------------------------+
| Loops in one session | (in hours, days)
+---+-----------------------------------------+
| (Posterity - Team)
+---+--- Repository --------------------------+
| Sessions in one repository | (in months, years)
+---------------------------------------------+
Dark Friction
The concluding SPIRIT chapter will tie us back to the cost of stewardship - an alternate framing being friction - and towards the dark software factory paradigm. 😉
AI disclaimer: Have bounced ideas off, so while text is still my own, the way I arrived at some of them may contain minor influences from such interactions.