Definition
Context starvation is a failure mode where an agent is missing the constraints, references, or history it needs to do correct work—so it fills the gaps with plausible guesses.
This is not “the model is dumb.” It’s “the system fed it a thin map.”
Two flavors
Accidental starvation (bad)
- The agent can’t see relevant files.
- It doesn’t know invariants (“don’t touch archived docs”).
- It’s unaware of prior decisions (ADRs, style rules, tooling versions).
Result: confident changes that don’t fit the system.
Intentional starvation (sometimes good)
Sometimes you want to constrain context:
- to reduce leakage risk,
- to keep a task tightly scoped,
- to prevent “helpful” refactors.
The difference is whether the agent still has the minimum necessary constraints.
Symptoms
- “I updated X” but touched unrelated files
- Reinventing existing utilities (“I wrote a new validator”) because it didn’t find the old one
- Breaking conventions (file names, schema fields)
- High narrative output, low artifact output (process confabulation)
Mitigations
- Start with an explore agent to produce pointers and constraints.
- Use schemas and linters to push work into validated lanes (schema gravity).
- Require evidence at the trust boundary.
- Keep tasks small enough that the needed context fits.
Practical rule
If an agent can’t cite the file/function it is changing and why, you are starving it (or it’s guessing).